Master Art of Translation | The Atmospheric Reanalysis, GRIB2 & CF-netCDF Translation System — How Forecast Times, Accumulations, Ensembles and Grid Metadata Keep Their Meaning

How do you translate GRIB2 weather files, netCDF climate data, ERA5 reanalysis variables and forecast terminology without changing the science? The answer begins with the meaning of every field, not the spelling of a column heading. People searching for GRIB2 translation, CF-netCDF metadata translation, ERA5 precipitation units, forecast reference time versus valid time, weather-model data localization, or multilingual climate-data documentation face a shared problem: a technically plausible translation can quietly change when, where, how or over what interval a number is valid. A sentence that reports “precipitation at 18:00” could mean an instantaneous rate, an hour ending at 18:00, a six-hour forecast accumulation, or a monthly diagnostic. These are different claims even when the numeric token is identical.

A reliable atmospheric reanalysis and weather forecast data translation therefore preserves the original record’s identity: producing centre, dataset and version, variable definition, native units, spatial grid, vertical reference, forecast initialization, lead time, validity time, temporal processing, ensemble membership and any missing-data or quality flags. GRIB2 is a coded meteorological message format; netCDF is a data container; CF metadata conventions describe the science represented by variables and coordinates. They are related but not interchangeable. The question is not whether a translator can replace “wind” with a term in another language. It is whether a competent reader or machine can reconstruct the same physical quantity with the same support and uncertainty.

This world-facing Master Art of Translation guide to GRIB2, CF-netCDF, ERA5 and ensemble forecast interpretation provides a practical system for translators, researchers, data engineers, editors, climate communicators, technical writers and reviewers. It shows how to translate forecast-step labels, accumulated rainfall, Kelvin and Celsius temperatures, Pa and hPa pressure, wind components, cell bounds, WMO code tables, CF standard names, time coordinates, grid-cell means, ensemble spread, revisions and provenance. It then supplies a repeatable evidence-preserving workflow, failure diagnostics, independently checkable hypothetical records, transfer exercises, release criteria and routes to related language learning. Its purpose is translation of documentation and meaning, not silent modification of machine-readable scientific fields.

Architecture record: EDKSG-TRANS-MASTER-WORLD-790. Scope: machine-readable atmospheric and climate evidence, multilingual explanatory prose, metadata interpretation and quality checking. This is a specialist child of Master Art of Translation | The Complete System for Moving Meaning Between Languages; it is not a replacement for that root or for existing vocabulary and English-learning curricula.

Reader routes: choose the question you need to answer

• “What does this GRIB2 forecast hour represent?” Start with reference time, lead-time units, product definition and interval processing.
• “How do I translate an ERA5 precipitation value?” Start with dataset/product identity, native units, accumulation window and valid time; do not multiply before documenting the source.
• “Why is my netCDF variable ambiguous?” Start with CF standard_name, units, coordinates, bounds, cell_methods and history.
• “Can I compare this field to a weather station?” Start with grid-cell support, observation exposure, height, representativeness and aggregation.
• “How can I publish a multilingual climate chart?” Start with axis semantics, ensemble labels, uncertainty explanations, source version, legends and unaltered identifiers.
• “What review prevents errors?” Use the twenty-eight release gates and worked diagnostic library below.

The search intent: translators need to preserve a measurement, not merely a noun

A search for “translate weather forecast” can mean public-language localization, a scientific dataset description, an operational bulletin, a software user-interface string, a map legend, a research methods section, or the metadata attached to a multidimensional array. These tasks have different failure consequences. Ordinary prose can usually be reorganized for readability. A field identifier such as a CF standard name, a versioned WMO code, or a coordinate-system attribute must generally remain machine-stable, with any translated explanation stored separately. Editing both indiscriminately may break software or destroy comparability.

The scientific source may be an analysis, a retrospective reanalysis, an operational forecast, an ensemble product, a climatological normal, an anomaly, or a user-generated aggregation. None of these labels is an interchangeable synonym for “weather data.” Analysis normally describes an estimated state at an analysis time; a forecast predicts a future valid time relative to a reference or initialization time; reanalysis uses a consistent assimilation framework to estimate past states and evolves across releases. A translator must ask which source operation produced the number before choosing natural-language verbs such as observed, measured, estimated, predicted or simulated.

Readers also search for how to convert GRIB2 to netCDF, but file conversion and linguistic translation are separate processes. Conversion may repack arrays, reorganize dimensions and map metadata. Translation can add localized variable descriptions and user guidance. Both require validation, yet neither authorizes changing the meaning of units, ensembles, accumulation periods or fill values. A successful pipeline preserves source data and creates a traceable representation of meaning; it does not treat a CSV header or map legend as the entire scientific record.

The hidden problem: weather data has several simultaneous clocks, spaces and authorities

A public forecast can look straightforward: a time, an icon and a number. Underneath, the number can belong to a particular model run, analysis cycle, forecast lead, ensemble member, vertical surface, geographic cell, temporal interval and processing method. The same field can appear in a GRIB2 product, a netCDF archive, a web service and a plotted image, each with different display conventions. Once translated into a sentence, most of those dimensions disappear unless the author deliberately preserves them.

Consider a hypothetical field whose reference is 2026-06-10 00:00 UTC, whose step is +18 hours and whose value represents precipitation accumulated over forecast hours 12–18. The field is valid at 18:00 UTC but describes an interval, not a single instant. Rewriting it as “rainfall measured at six in the evening” substitutes measurement for modelled quantity, hides the interval and suggests a local clock. Rewriting it as “rainfall over the previous eighteen hours” also fails because the stored interval is six hours. The safe translation identifies the run and window explicitly and explains that the total is forecast, not observed.

The second hidden problem is authority. WMO code tables, the originating centre’s documentation, dataset-specific processing notes and CF metadata can each answer different questions. A netCDF conversion may carry a useful CF standard_name and a human-readable long_name, but dataset-specific interpretation is still needed for accumulation windows or forecast conventions. Reviewers must preserve both the machine encoding and the supplier’s documented semantics rather than selecting a single attractive phrase.

High-authority reference frame, verified for the 2026 publication

The WMO Manual on Codes Volume I.2 machine-readable codes identifies the 1 June 2026 implementation set as GRIB edition 2 version 37 and BUFR edition 4 version 46. This article refers to those published versions, not to a draft planned for later implementation. The WMO Codes Registry supplies structured code-table identifiers and status information. Code numbers have meaning only with their discipline, category, table, local-table context and applicable version where relevant.

The CF Metadata Conventions 1.13 were published on 17 December 2025. They distinguish standard_name, units, coordinate bounds and cell_methods, allowing a variable to describe its physical quantity and the intervals or cells it represents. The ECMWF ERA5 product overview describes an hourly global reanalysis from 1940 onward, and the Copernicus reanalysis overview explains the distinction between reanalysis, estimates of uncertainty and near-real-time provisional releases. These are sources for terminology and method, not substitute metadata for an arbitrary downloaded file.

For assessment of translated outputs, ISO 5060:2024 outlines human evaluation of translation outputs with agreed error categories, sampling and scoring. Its scope is translation evaluation; it is not a meteorological data standard, a substitute for unit tests, or a claim that a dataset is scientifically validated. Any example in this guide not explicitly identified as a documented external fact is a constructed training record. Numbers are illustrative, never live forecasts, warnings, archived measurements or official product codes.

The mechanism: a semantic envelope around every translated value

Before translation, define an envelope containing the following: dataset identifier, producing organization, product subtype, generation or edition, processing level, variable identifier, canonical definition, original data type, unit, vertical and horizontal reference, time reference, validity, interval support, forecast lead, ensemble descriptor, precision, missing-data rules, source or licence, and revision history. Some variables do not use every field. The reason for an explicit envelope is that removing an unused field is different from omitting a field whose value is unknown.

The envelope lets reviewers test reversibility. Given only the target-language description and the protected original identifiers, can a domain reader reconstruct what the value actually means? This reconstruction is stronger than surface fluency. A beautiful translation is inadequate if it converts a grid-box water equivalent into a gauge measurement, a forecast into an observation, a pressure level into an altitude, or an ensemble spread into a range of certainties.

Make two tracks: one immutable source track containing raw identifiers, scientific values and documented metadata, and one editable reader track containing translated explanatory language. Map each translated sentence back to the first track. A multilingual interface may present accessible phrases in any writing system while retaining the raw variable, unit, coordinate and model run unchanged. This principle allows human localization without corrupting software-facing data.

The 80-part technical mechanism atlas: interpret first, translate second

The following 100 mechanisms are organized as a dependency map rather than as a dictionary. Each begins with a physical or information-system principle, names a predictable translation failure, presents a constructed diagnostic and gives a reconstruction test. A reviewer can use one entry in isolation, but the stronger use is to chain the entries: a precipitation field may need the forecast-reference clock, time bounds, water-equivalent units, spatial support and provenance together.

01. Analysis field

First principle. An estimated atmospheric state at an analysis time, often informed by data assimilation. This is a semantic constraint, not merely a preferred term. Define the origin of authority before choosing the translated noun. Scientific identifiers, value definitions and time support form the invariant core; natural-language ordering and register are flexible. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Replacing estimated with directly observed erases modelling and analysis context. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. Hypothetical temperature analysis of 289 K at 00 UTC is not itself an individual thermometer measurement. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Write analysis estimate, state time, and keep its provenance separate from station observations. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

02. Operational forecast

First principle. A model prediction tied to an initialization and a lead time. This is a semantic constraint, not merely a preferred term. Ask which property is asserted, which quantities are merely metadata and which limitations prevent stronger conclusions. Words such as measured, estimated, averaged, predicted and derived are not substitutes. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Calling a forecast an observation falsely upgrades evidence. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A 12-hour forecast initialized at 00 UTC can be valid at 12 UTC although it was produced earlier. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Pair initialization with valid time and retain forecast designation. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

03. Reanalysis product

First principle. A retrospective, model-assisted reconstruction using observations and a consistent production system. This is a semantic constraint, not merely a preferred term. Use two versions of every short example: a readable sentence for the intended audience and a machine-facing description retaining raw identifiers and scope. Review both against the same source. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Calling reanalysis raw historical weather observations hides assimilation and uncertainty. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. An ERA5 hourly value estimates a past state over a grid rather than reproducing one station record. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Retain product name, release and reanalysis classification. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

04. Hindcast

First principle. A retrospective model prediction or simulation used to evaluate behaviour over historical periods. This is a semantic constraint, not merely a preferred term. The decisive test is not whether the sentence is elegant. It is whether a colleague who did not inspect the source file could reconstruct the same value class, location, period and method without importing assumptions. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Confusing hindcast and reanalysis loses how observations were used. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A seasonal forecast model run retrospectively for old years may not assimilate observations like a reanalysis. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Document simulation design and whether a run was initialized with historical inputs. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

05. Nowcast

First principle. A very short-range estimate or forecast with product-specific methodology. This is a semantic constraint, not merely a preferred term. Define the origin of authority before choosing the translated noun. Scientific identifiers, value definitions and time support form the invariant core; natural-language ordering and register are flexible. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Treating nowcast as a verified current observation overclaims certainty. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. An extrapolated precipitation radar field may represent the next short interval rather than an instrument observation. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Preserve product label and latency or forecast horizon. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

06. Climate projection

First principle. A conditional future model result based on scenarios and assumptions. This is a semantic constraint, not merely a preferred term. Ask which property is asserted, which quantities are merely metadata and which limitations prevent stronger conclusions. Words such as measured, estimated, averaged, predicted and derived are not substitutes. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Calling a scenario-dependent projection a weather forecast suggests different predictive certainty. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. An ensemble under one emissions pathway is not a prediction of one afternoon’s weather. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Preserve scenario, model, period and conditional language. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

07. Climatological normal

First principle. A statistic calculated for a specified multi-year reference interval. This is a semantic constraint, not merely a preferred term. Use two versions of every short example: a readable sentence for the intended audience and a machine-facing description retaining raw identifiers and scope. Review both against the same source. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Using normal to imply natural or desirable changes the claim. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A monthly temperature normal over a 30-year baseline is a reference statistic, not a physiological ideal. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. State statistic and baseline dates. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

08. Climate anomaly

First principle. A departure relative to a stated reference climatology or baseline. This is a semantic constraint, not merely a preferred term. The decisive test is not whether the sentence is elegant. It is whether a colleague who did not inspect the source file could reconstruct the same value class, location, period and method without importing assumptions. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Printing the departure without its baseline makes comparison ambiguous. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. An anomaly of plus 1.2 K requires the reader to know the reference period and statistic. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Translate anomaly as difference from the named baseline. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

09. Dataset edition

First principle. The identity of a delivered product and its particular published version. This is a semantic constraint, not merely a preferred term. Define the origin of authority before choosing the translated noun. Scientific identifiers, value definitions and time support form the invariant core; natural-language ordering and register are flexible. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Mixing values or descriptions from two releases can manufacture apparent trends. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. Two releases may share variable names while differing in bias correction or timeliness. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Preserve version, release date and retrieval date. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

10. Provisional release

First principle. A near-real-time release with a documented path to later replacement or adjustment. This is a semantic constraint, not merely a preferred term. Ask which property is asserted, which quantities are merely metadata and which limitations prevent stronger conclusions. Words such as measured, estimated, averaged, predicted and derived are not substitutes. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Suppressing provisional status portrays revision-prone data as final. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A recent reanalysis product labelled provisional must not be described as a frozen historical record. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Keep provisional qualifier and the expected revision pathway without inventing dates. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

11. Analysis reference time

First principle. The timestamp associated with an analysis or initialization. This is a semantic constraint, not merely a preferred term. Use two versions of every short example: a readable sentence for the intended audience and a machine-facing description retaining raw identifiers and scope. Review both against the same source. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Treating reference time as the time represented by any forecast step misplaces events. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A run reference 2026-04-08T00Z plus a 24-hour step has a different valid time. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Store both clocks and their relation. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

12. Forecast validity time

First principle. The time at which an instantaneous forecast applies or an interval field is indexed. This is a semantic constraint, not merely a preferred term. The decisive test is not whether the sentence is elegant. It is whether a colleague who did not inspect the source file could reconstruct the same value class, location, period and method without importing assumptions. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Describing valid time as the start of any accumulation creates a shifted timeline. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. An accumulated field valid at 18 UTC may summarize hours 12 through 18. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Display validity and interval bounds separately. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

13. Forecast lead

First principle. Elapsed forecast time from the reference cycle, with units defined by the product. This is a semantic constraint, not merely a preferred term. Define the origin of authority before choosing the translated noun. Scientific identifiers, value definitions and time support form the invariant core; natural-language ordering and register are flexible. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Assuming a bare 12 always means twelve hours ignores code-table time-unit choices. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. The same lead duration may be encoded in minutes or hours. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Decode unit and numerical lead before prose translation. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

14. Production time

First principle. When a file was generated, assembled, or delivered. This is a semantic constraint, not merely a preferred term. Ask which property is asserted, which quantities are merely metadata and which limitations prevent stronger conclusions. Words such as measured, estimated, averaged, predicted and derived are not substitutes. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Confusing file creation with phenomenon time makes stale products appear current. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A forecast copied the next day retains its original initialization and validity. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Keep generation timestamp as distribution metadata. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

15. Time-zone indication

First principle. A declared convention for interpreting timestamps. This is a semantic constraint, not merely a preferred term. Use two versions of every short example: a readable sentence for the intended audience and a machine-facing description retaining raw identifiers and scope. Review both against the same source. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Adding an unqualified local-time label shifts the temporal referent. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. An ISO timestamp ending in Z is UTC, not a clock reading in every reader’s locale. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Preserve ISO timestamp and display locale conversion as an additional view. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

16. Calendar attribute

First principle. The calendar model used by a netCDF time coordinate. This is a semantic constraint, not merely a preferred term. The decisive test is not whether the sentence is elegant. It is whether a colleague who did not inspect the source file could reconstruct the same value class, location, period and method without importing assumptions. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Applying Gregorian arithmetic to a 360_day model calendar produces false dates. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A model month with thirty days in a 360_day calendar is not necessarily a real calendar month. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Keep calendar attribute and describe model-time conventions. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

17. Time coordinate units

First principle. The numeric epoch and duration units used for time values. This is a semantic constraint, not merely a preferred term. Define the origin of authority before choosing the translated noun. Scientific identifiers, value definitions and time support form the invariant core; natural-language ordering and register are flexible. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Reading hours since a base date as an absolute timestamp produces nonsense. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A time coordinate 18 with units hours since a UTC date encodes an offset, not hour-of-day by itself. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Decode units and epoch together. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

18. Instantaneous sample

First principle. A value representative of a particular time or instant. This is a semantic constraint, not merely a preferred term. Ask which property is asserted, which quantities are merely metadata and which limitations prevent stronger conclusions. Words such as measured, estimated, averaged, predicted and derived are not substitutes. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Calling a sample an hourly mean claims an averaging operation that never occurred. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. Surface pressure at 09 UTC is not automatically an average between 08 and 09 UTC. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Use at time unless interval processing is documented. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

19. Temporal mean

First principle. A statistic over a declared interval with beginning and end. This is a semantic constraint, not merely a preferred term. Use two versions of every short example: a readable sentence for the intended audience and a machine-facing description retaining raw identifiers and scope. Review both against the same source. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Presenting mean temperature as a point observation conceals its time support. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A daily mean from 24 hourly fields depends on which 24 fields and time basis were selected. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Preserve averaging method and interval bounds. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

20. Temporal sum

First principle. A quantity accumulated or summed across a time interval. This is a semantic constraint, not merely a preferred term. The decisive test is not whether the sentence is elegant. It is whether a colleague who did not inspect the source file could reconstruct the same value class, location, period and method without importing assumptions. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Confusing total depth with rate can inflate totals by hours or seconds. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A six-hour water equivalent of 0.012 m represents 12 mm over that interval, not 12 mm per hour. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Report total, duration and interval endpoints. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

21. Temporal maximum

First principle. The greatest reported value within a stated window. This is a semantic constraint, not merely a preferred term. Define the origin of authority before choosing the translated noun. Scientific identifiers, value definitions and time support form the invariant core; natural-language ordering and register are flexible. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Writing maximum as instantaneous value misrepresents duration. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A maximum wind gust over three hours cannot be assigned to the window endpoint without source evidence. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Preserve window and method. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

22. Temporal minimum

First principle. The lowest reported value within a stated window. This is a semantic constraint, not merely a preferred term. Ask which property is asserted, which quantities are merely metadata and which limitations prevent stronger conclusions. Words such as measured, estimated, averaged, predicted and derived are not substitutes. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Inferring the exact occurrence instant from window validity adds unsupported precision. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A six-hour minimum temperature indexed by 06 UTC may have occurred earlier. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Keep window and avoid inventing event time. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

23. Rolling window

First principle. A statistic calculated for a window that moves between successive outputs. This is a semantic constraint, not merely a preferred term. Use two versions of every short example: a readable sentence for the intended audience and a machine-facing description retaining raw identifiers and scope. Review both against the same source. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Summing overlapping rolling windows double-counts the same underlying hours. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. Consecutive twelve-hour rainfall totals overlapping by eleven hours cannot be added as independent totals. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Mark window length, movement and overlap. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

24. Accumulation reset

First principle. The point at which a run or product resets a cumulative quantity. This is a semantic constraint, not merely a preferred term. The decisive test is not whether the sentence is elegant. It is whether a colleague who did not inspect the source file could reconstruct the same value class, location, period and method without importing assumptions. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Differencing across a reset may fabricate a negative rainfall amount. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. The zero-hour value and next cycle’s accumulation are not necessarily continuous. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Check cycle boundaries before deriving increments. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

25. Accumulation beginning

First principle. The first time in the temporal support of an accumulated field. This is a semantic constraint, not merely a preferred term. Define the origin of authority before choosing the translated noun. Scientific identifiers, value definitions and time support form the invariant core; natural-language ordering and register are flexible. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Assuming beginning equals initialization loses partial-window products. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A 12 to 18 hour forecast sum begins at hour 12, not at forecast hour zero. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Record beginning lead and ending lead. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

26. Accumulation end

First principle. The endpoint that indexes a completed accumulation. This is a semantic constraint, not merely a preferred term. Ask which property is asserted, which quantities are merely metadata and which limitations prevent stronger conclusions. Words such as measured, estimated, averaged, predicted and derived are not substitutes. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Treating the endpoint as a single-instant observation changes the value class. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A forecast labelled valid 18 UTC may cover the preceding six hours. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Explain end time and duration. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

27. Forecast step range

First principle. Two lead-time endpoints used to identify the supported interval. This is a semantic constraint, not merely a preferred term. Use two versions of every short example: a readable sentence for the intended audience and a machine-facing description retaining raw identifiers and scope. Review both against the same source. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Translating only the latter number hides whether the statistic was accumulated from zero. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. The intervals 0 to 18 and 12 to 18 both end at hour 18 but differ materially. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Retain both step boundaries. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

28. Accumulation versus rate

First principle. The distinction between total over a window and change per unit time. This is a semantic constraint, not merely a preferred term. The decisive test is not whether the sentence is elegant. It is whether a colleague who did not inspect the source file could reconstruct the same value class, location, period and method without importing assumptions. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Adding rates as if they were totals ignores integration. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A precipitation rate of 0.001 kg m−2 s−1 must not be described as one millimetre without a duration. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. State units, integration window and conversion assumptions. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

29. Kelvin temperature

First principle. An absolute thermodynamic temperature with symbol K. This is a semantic constraint, not merely a preferred term. Define the origin of authority before choosing the translated noun. Scientific identifiers, value definitions and time support form the invariant core; natural-language ordering and register are flexible. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Appending degrees or treating a change like an absolute temperature causes errors. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A temperature of 293.15 K is 20 degrees Celsius, whereas a difference of 1 K equals a difference of 1 Celsius degree. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Preserve absolute versus difference semantics. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

30. Celsius display

First principle. A derived presentation of temperature from a compatible absolute source. This is a semantic constraint, not merely a preferred term. Ask which property is asserted, which quantities are merely metadata and which limitations prevent stronger conclusions. Words such as measured, estimated, averaged, predicted and derived are not substitutes. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Changing values without labelling the converted units breaks reproducibility. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A map legend converted from Kelvin to Celsius must document subtracting 273.15 from absolute temperature. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Keep native array and make display transform explicit. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

31. Surface pressure

First principle. Atmospheric pressure at the model surface or terrain height. This is a semantic constraint, not merely a preferred term. Use two versions of every short example: a readable sentence for the intended audience and a machine-facing description retaining raw identifiers and scope. Review both against the same source. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Calling surface pressure sea-level pressure disregards elevation and reduction method. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A high-elevation grid cell can have lower surface pressure without a storm. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Retain surface reference and Pa or hPa unit. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

32. Mean sea-level pressure

First principle. Pressure adjusted to a reference at mean sea level using a defined method. This is a semantic constraint, not merely a preferred term. The decisive test is not whether the sentence is elegant. It is whether a colleague who did not inspect the source file could reconstruct the same value class, location, period and method without importing assumptions. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Treating it as measured at ground level creates a false instrument comparison. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. An MSLP contour differs semantically from a station’s actual surface pressure. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Retain reduction concept. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

33. Geopotential

First principle. Potential energy per unit mass associated with gravity in meteorology. This is a semantic constraint, not merely a preferred term. Define the origin of authority before choosing the translated noun. Scientific identifiers, value definitions and time support form the invariant core; natural-language ordering and register are flexible. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Equating geopotential with geometric altitude without transformation can misstate vertical positions. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A 500 hPa geopotential field has unit m2 s−2 in some products and may be displayed as geopotential height after division by gravity. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Document source parameter and any converted product. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

34. Pressure level

First principle. An isobaric vertical surface identified by pressure rather than geometric height. This is a semantic constraint, not merely a preferred term. Ask which property is asserted, which quantities are merely metadata and which limitations prevent stronger conclusions. Words such as measured, estimated, averaged, predicted and derived are not substitutes. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Interpreting 500 hPa as 500 metres is a categorical unit error. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. Temperature at 500 hPa represents the specified pressure surface which varies in height. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Preserve hPa label and variable’s vertical coordinate. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

35. Hybrid model level

First principle. A terrain-following vertical coordinate defined by formula terms. This is a semantic constraint, not merely a preferred term. Use two versions of every short example: a readable sentence for the intended audience and a machine-facing description retaining raw identifiers and scope. Review both against the same source. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Translating a model level number into metres invents altitude. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. The geometric height of level 40 depends on pressure and surface terms, not the ordinal 40 alone. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Retain formula_terms and model-level identity. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

36. Height above ground

First principle. An explicit vertical displacement relative to ground or a specified surface. This is a semantic constraint, not merely a preferred term. The decisive test is not whether the sentence is elegant. It is whether a colleague who did not inspect the source file could reconstruct the same value class, location, period and method without importing assumptions. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Omitting the reference turns 2 m air temperature into an undefined vertical observation. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A 10 m wind is not a wind measured at arbitrary rooftop height. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Preserve height and datum. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

37. Grid-cell mean

First principle. A quantity representative of a spatial cell or support, not a specific point sensor. This is a semantic constraint, not merely a preferred term. Define the origin of authority before choosing the translated noun. Scientific identifiers, value definitions and time support form the invariant core; natural-language ordering and register are flexible. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Writing at a named city can imply exact local measurement. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A coarse grid covers variable terrain and cannot promise a thermometer reading at one address. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Use over the grid cell and note spatial resolution. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

38. Grid-point value

First principle. A quantity evaluated at a discrete grid coordinate with model support. This is a semantic constraint, not merely a preferred term. Ask which property is asserted, which quantities are merely metadata and which limitations prevent stronger conclusions. Words such as measured, estimated, averaged, predicted and derived are not substitutes. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Describing one point as a regional average can change representativeness. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A point interpolation and cell mean have different supports even at similar displayed coordinates. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Record interpolation and grid definition. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

39. Latitude ordering

First principle. Whether array rows run north to south or south to north. This is a semantic constraint, not merely a preferred term. Use two versions of every short example: a readable sentence for the intended audience and a machine-facing description retaining raw identifiers and scope. Review both against the same source. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Reversing row order while keeping latitude labels unchanged flips a map vertically. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A stored latitude sequence 90, 89, 88 is descending and must match array indices. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Validate coordinate monotonicity and bounds. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

40. Longitude convention

First principle. Whether longitudes are represented as zero to 360 or minus 180 to 180. This is a semantic constraint, not merely a preferred term. The decisive test is not whether the sentence is elegant. It is whether a colleague who did not inspect the source file could reconstruct the same value class, location, period and method without importing assumptions. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Treating an equivalent wrapped longitude as a different physical position duplicates points. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. Longitude 350 degrees east is equivalent to minus 10 degrees, not a site ten degrees east. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Normalize only for display with reversible mapping. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

41. Rotated grid

First principle. A coordinate grid whose axes differ from ordinary geodetic latitude and longitude. This is a semantic constraint, not merely a preferred term. Define the origin of authority before choosing the translated noun. Scientific identifiers, value definitions and time support form the invariant core; natural-language ordering and register are flexible. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Labelling rotated coordinates directly as Earth latitude and longitude shifts locations. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A rotated-pole grid needs its grid mapping to locate a cell correctly. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Preserve rotated coordinate definition. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

42. Projected grid

First principle. A coordinate system defined by map projection and its parameters. This is a semantic constraint, not merely a preferred term. Ask which property is asserted, which quantities are merely metadata and which limitations prevent stronger conclusions. Words such as measured, estimated, averaged, predicted and derived are not substitutes. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Relabelling projected x/y as latitude/longitude destroys geographic meaning. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A Lambert conformal grid requires its standard parallels and projection origin for reproducibility. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Carry projection metadata and coordinates. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

43. Grid mapping

First principle. CF metadata connecting native coordinates to a coordinate reference system. This is a semantic constraint, not merely a preferred term. Use two versions of every short example: a readable sentence for the intended audience and a machine-facing description retaining raw identifiers and scope. Review both against the same source. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Deleting the mapping can make values appear in a different region. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. An image may look visually plausible while the underlying transform has vanished. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Retain grid_mapping and a separately localized description. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

44. Cell bounds

First principle. Coordinates delineating a cell’s interval or spatial area. This is a semantic constraint, not merely a preferred term. The decisive test is not whether the sentence is elegant. It is whether a colleague who did not inspect the source file could reconstruct the same value class, location, period and method without importing assumptions. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Assuming a point coordinate always means the centre of a rectangular cell introduces unsupported geometry. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. CF bounds can locate temporal or spatial support more precisely than coordinate labels alone. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Preserve bounds and do not invent centroids. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

45. Cell area

First principle. Area used for spatial weighting or integration. This is a semantic constraint, not merely a preferred term. Define the origin of authority before choosing the translated noun. Scientific identifiers, value definitions and time support form the invariant core; natural-language ordering and register are flexible. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Averaging unequal-area cells without weights can bias a global statistic. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. One degree of longitude spans different physical distances at different latitudes. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Retain cell_measures or appropriate area weights. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

46. Land-sea mask

First principle. A field identifying land/ocean fraction or classification in a cell. This is a semantic constraint, not merely a preferred term. Ask which property is asserted, which quantities are merely metadata and which limitations prevent stronger conclusions. Words such as measured, estimated, averaged, predicted and derived are not substitutes. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Treating partial land fraction as an absolute coastline invents resolution. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A coastal cell with land fraction 0.4 contains both surface types, not a perfectly mapped shoreline. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Explain fractional or categorical mask and grid support. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

47. Oroography

First principle. Representation of surface elevation in a model grid. This is a semantic constraint, not merely a preferred term. Use two versions of every short example: a readable sentence for the intended audience and a machine-facing description retaining raw identifiers and scope. Review both against the same source. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Describing model surface altitude as precise local topography falsely sharpens the data. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A grid elevation is an aggregate or model-representative height. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Label terrain representation and source resolution. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

48. Interpolation method

First principle. A transformation estimating values onto new coordinates. This is a semantic constraint, not merely a preferred term. The decisive test is not whether the sentence is elegant. It is whether a colleague who did not inspect the source file could reconstruct the same value class, location, period and method without importing assumptions. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Calling regridded numbers native measurements conceals a derived calculation. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. Nearest neighbour and bilinear interpolation can produce different extreme values. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Record source grid, target grid and method. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

49. Regridding conservativeness

First principle. Whether a spatial transform conserves an extensive quantity under stated assumptions. This is a semantic constraint, not merely a preferred term. Define the origin of authority before choosing the translated noun. Scientific identifiers, value definitions and time support form the invariant core; natural-language ordering and register are flexible. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Applying a generic interpolator to area totals may change integrated quantities. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. Conservative remapping is a property to verify, not a guarantee from a translated label. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Keep method and check integrated totals. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

50. Wind U component

First principle. The eastward or grid-relative component of wind velocity. This is a semantic constraint, not merely a preferred term. Ask which property is asserted, which quantities are merely metadata and which limitations prevent stronger conclusions. Words such as measured, estimated, averaged, predicted and derived are not substitutes. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Rendering U as wind speed ignores vector direction and possible grid rotation. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. U equals minus 5 m/s describes a westward component in an east-positive coordinate system. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. State component orientation and sign. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

51. Wind V component

First principle. The northward or grid-relative wind component. This is a semantic constraint, not merely a preferred term. Use two versions of every short example: a readable sentence for the intended audience and a machine-facing description retaining raw identifiers and scope. Review both against the same source. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Replacing V with a plain speed loses vector orientation. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A V value of plus 4 m/s is a component, not proof of a four metre per second total wind. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Keep coordinate basis and component label. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

52. Wind speed magnitude

First principle. The magnitude calculated from vector components or supplied by a model. This is a semantic constraint, not merely a preferred term. The decisive test is not whether the sentence is elegant. It is whether a colleague who did not inspect the source file could reconstruct the same value class, location, period and method without importing assumptions. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Taking arithmetic means of magnitudes as vector-mean wind changes statistics. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A speed of 10 m/s can arise from different directions and cannot reveal its own components. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Preserve magnitude and aggregation method. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

53. Wind direction convention

First principle. The convention for direction from which or toward which wind travels. This is a semantic constraint, not merely a preferred term. Define the origin of authority before choosing the translated noun. Scientific identifiers, value definitions and time support form the invariant core; natural-language ordering and register are flexible. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Flipping from-direction to to-direction changes meaning by 180 degrees. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A meteorological northerly wind comes from the north, not travels toward it. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Declare convention and angle reference. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

54. Wind gust

First principle. A short-duration peak under a model or observational definition. This is a semantic constraint, not merely a preferred term. Ask which property is asserted, which quantities are merely metadata and which limitations prevent stronger conclusions. Words such as measured, estimated, averaged, predicted and derived are not substitutes. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Treating a gust as mean wind speed exaggerates sustained conditions. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A maximum gust of 22 m/s cannot be read as a ten-minute average without evidence. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Preserve peak definition and window. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

55. Dew point

First principle. Temperature at which air reaches saturation for a given moisture state. This is a semantic constraint, not merely a preferred term. Use two versions of every short example: a readable sentence for the intended audience and a machine-facing description retaining raw identifiers and scope. Review both against the same source. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Mistaking dew point for air temperature changes humidity interpretation. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. Air temperature 28 C and dew point 21 C are different fields even when a chart shows temperatures. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Retain physical definition. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

56. Specific humidity

First principle. Mass of water vapour per mass of moist air. This is a semantic constraint, not merely a preferred term. The decisive test is not whether the sentence is elegant. It is whether a colleague who did not inspect the source file could reconstruct the same value class, location, period and method without importing assumptions. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Confusing specific humidity and mixing ratio changes denominator. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. 0.010 kg/kg of moist air does not directly equal a percent relative humidity. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Preserve exact property and unit. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

57. Relative humidity

First principle. Vapour pressure relative to saturation under a defined temperature basis. This is a semantic constraint, not merely a preferred term. Define the origin of authority before choosing the translated noun. Scientific identifiers, value definitions and time support form the invariant core; natural-language ordering and register are flexible. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Treating percent relative humidity as mass percentage creates a false material balance. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. 80 percent RH is not 80 percent water by mass. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Keep saturation-based definition and percent unit. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

58. Total cloud cover

First principle. Fraction of sky or grid area represented by cloud cover under a product definition. This is a semantic constraint, not merely a preferred term. Ask which property is asserted, which quantities are merely metadata and which limitations prevent stronger conclusions. Words such as measured, estimated, averaged, predicted and derived are not substitutes. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Treating fractional cover as cloud thickness or rain probability changes variables. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A cover fraction of 0.75 means 75 percent under the product’s definition, not a 75 percent precipitation chance. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Label coverage and support. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

59. Convective precipitation

First principle. Precipitation parameter attributed to a model’s convective scheme. This is a semantic constraint, not merely a preferred term. Use two versions of every short example: a readable sentence for the intended audience and a machine-facing description retaining raw identifiers and scope. Review both against the same source. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Translating the category as all thunderstorm rainfall may overstate physical attribution. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A model separates convective and large-scale contributions by parameterization, which differs by system. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Retain parameter distinction. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

60. Large-scale precipitation

First principle. Precipitation attributed to resolved cloud or large-scale processes in the model. This is a semantic constraint, not merely a preferred term. The decisive test is not whether the sentence is elegant. It is whether a colleague who did not inspect the source file could reconstruct the same value class, location, period and method without importing assumptions. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Equating large-scale with wide geographic extent is too literal. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A labelled component is defined by scheme, not a public claim about affected area. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Explain model process classification. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

61. Total precipitation

First principle. Water-equivalent precipitation amount over a defined interval. This is a semantic constraint, not merely a preferred term. Define the origin of authority before choosing the translated noun. Scientific identifiers, value definitions and time support form the invariant core; natural-language ordering and register are flexible. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Replacing it with rain amount discards snow or other included frozen precipitation. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A product total of 0.008 m water equivalent represents eight millimetres for its declared window. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Keep inclusions, units and accumulation bounds. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

62. Snowfall water equivalent

First principle. Liquid-water equivalent corresponding to a snow quantity. This is a semantic constraint, not merely a preferred term. Ask which property is asserted, which quantities are merely metadata and which limitations prevent stronger conclusions. Words such as measured, estimated, averaged, predicted and derived are not substitutes. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Equating it directly with snow depth ignores density. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. Ten millimetres water equivalent is not necessarily ten millimetres snow depth. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Identify water-equivalent basis. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

63. Snow depth

First principle. The modelled or measured vertical depth of snow cover. This is a semantic constraint, not merely a preferred term. Use two versions of every short example: a readable sentence for the intended audience and a machine-facing description retaining raw identifiers and scope. Review both against the same source. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Misreading depth as water equivalent breaks storage estimates. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A snowpack 0.30 m deep can have much smaller water equivalent depending on density. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Keep physical depth and unit. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

64. Surface runoff

First principle. Water that leaves the land surface as flow under a model’s water balance. This is a semantic constraint, not merely a preferred term. The decisive test is not whether the sentence is elegant. It is whether a colleague who did not inspect the source file could reconstruct the same value class, location, period and method without importing assumptions. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Calling it observed river discharge changes location and unit. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A runoff depth per grid area does not equal cubic metres per second at a gauge. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Report depth and spatial basis. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

65. Subsurface runoff

First principle. Modelled below-surface drainage component over a given time. This is a semantic constraint, not merely a preferred term. Define the origin of authority before choosing the translated noun. Scientific identifiers, value definitions and time support form the invariant core; natural-language ordering and register are flexible. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Calling all subsurface runoff groundwater recharge ignores routing and definitions. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A stored drainage depth may contribute to routing without equalling an aquifer recharge observation. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Preserve process and timescale. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

66. Soil moisture layer

First principle. Water content stored in a specified soil depth or model layer. This is a semantic constraint, not merely a preferred term. Ask which property is asserted, which quantities are merely metadata and which limitations prevent stronger conclusions. Words such as measured, estimated, averaged, predicted and derived are not substitutes. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Conflating volumetric water content with depth-integrated water storage changes units. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A value in m3/m3 is dimensionless fraction rather than millimetres of water. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Preserve layer bounds and water-content basis. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

67. Shortwave downward radiation

First principle. Solar-origin radiation incident on a surface, expressed as flux or accumulated energy. This is a semantic constraint, not merely a preferred term. Use two versions of every short example: a readable sentence for the intended audience and a machine-facing description retaining raw identifiers and scope. Review both against the same source. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Confusing J m−2 totals with W m−2 rates changes magnitude and duration. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. An accumulated surface energy of 3.6 MJ m−2 over one hour corresponds to a mean flux of 1000 W m−2 under exact constant-interval assumptions. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Report direction, unit and interval. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

68. Longwave radiation

First principle. Thermal infrared radiative flux or accumulated energy with directional specification. This is a semantic constraint, not merely a preferred term. The decisive test is not whether the sentence is elegant. It is whether a colleague who did not inspect the source file could reconstruct the same value class, location, period and method without importing assumptions. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Removing upward or downward can reverse an energy-budget interpretation. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. Downwelling and upwelling longwave terms are physically distinct. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Keep radiation band and direction. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

69. Sensible heat flux

First principle. Turbulent transfer of heat without phase-change latent energy. This is a semantic constraint, not merely a preferred term. Define the origin of authority before choosing the translated noun. Scientific identifiers, value definitions and time support form the invariant core; natural-language ordering and register are flexible. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Dropping sign convention can reverse surface heating interpretation. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A negative flux in one product may represent upward energy under a stated convention, while another uses opposite sign. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Check positive-direction rule. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

70. Latent heat flux

First principle. Energy associated with moisture phase change and turbulent exchange. This is a semantic constraint, not merely a preferred term. Ask which property is asserted, which quantities are merely metadata and which limitations prevent stronger conclusions. Words such as measured, estimated, averaged, predicted and derived are not substitutes. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Treating it as a measured evaporation depth without latent heat conversion is invalid. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. An energy flux needs duration and latent-heat assumptions to estimate water mass exchange. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Preserve energy form and sign. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

71. CAPE

First principle. Convective available potential energy of an atmospheric parcel under specified assumptions. This is a semantic constraint, not merely a preferred term. Use two versions of every short example: a readable sentence for the intended audience and a machine-facing description retaining raw identifiers and scope. Review both against the same source. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Calling CAPE a probability of storms is a category error. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A CAPE value in J/kg indicates parcel buoyancy energy, not a direct event probability. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Preserve diagnostic nature and units. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

72. Boundary-layer height

First principle. A model-derived estimate of atmospheric boundary-layer depth. This is a semantic constraint, not merely a preferred term. The decisive test is not whether the sentence is elegant. It is whether a colleague who did not inspect the source file could reconstruct the same value class, location, period and method without importing assumptions. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Reporting it as the altitude of every cloud base invents a relationship. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A mixed-layer depth can differ from cloud-base or inversion height. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Retain diagnostic definition. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

73. Visibility

First principle. A visibility diagnostic or observation with model-dependent definition. This is a semantic constraint, not merely a preferred term. Define the origin of authority before choosing the translated noun. Scientific identifiers, value definitions and time support form the invariant core; natural-language ordering and register are flexible. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Interpreting visual range as proof of clear air under all conditions overclaims. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A model field may use extinction estimates rather than direct human observation. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Specify method and unit. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

74. Ozone mixing ratio

First principle. An atmospheric composition quantity with a specific concentration or mixing-ratio unit. This is a semantic constraint, not merely a preferred term. Ask which property is asserted, which quantities are merely metadata and which limitations prevent stronger conclusions. Words such as measured, estimated, averaged, predicted and derived are not substitutes. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Treating all ozone fields as surface air quality changes altitude and relevance. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. Ozone at a pressure level differs from surface concentration used in exposure assessment. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Preserve vertical level and composition basis. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

75. Ensemble control

First principle. A designated member produced with the system’s control configuration. This is a semantic constraint, not merely a preferred term. Use two versions of every short example: a readable sentence for the intended audience and a machine-facing description retaining raw identifiers and scope. Review both against the same source. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Calling the control member the most likely outcome misstates ensemble interpretation. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. Control is a member label, not automatically a probability maximum. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Keep member identity and generation description. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

76. Perturbed member

First principle. An ensemble realization with an intentional perturbation in initial conditions or model representation. This is a semantic constraint, not merely a preferred term. The decisive test is not whether the sentence is elegant. It is whether a colleague who did not inspect the source file could reconstruct the same value class, location, period and method without importing assumptions. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Treating member number as ranking invents an order. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. Member 17 is an identifier, not seventeenth-best forecast. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Preserve member number without probability claims. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

77. Ensemble mean

First principle. A statistic across selected members at a specified time and variable. This is a semantic constraint, not merely a preferred term. Define the origin of authority before choosing the translated noun. Scientific identifiers, value definitions and time support form the invariant core; natural-language ordering and register are flexible. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Treating the mean as an actual member’s trajectory may create impossible combinations. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. An ensemble-mean rainfall amount is not necessarily realized by any one member. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. State aggregation set and statistic. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

78. Ensemble spread

First principle. A dispersion statistic whose definition may be standard deviation or another measure. This is a semantic constraint, not merely a preferred term. Ask which property is asserted, which quantities are merely metadata and which limitations prevent stronger conclusions. Words such as measured, estimated, averaged, predicted and derived are not substitutes. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Translating spread as error or confidence interval without method overstates inference. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A standard deviation of member temperatures is not an observational measurement uncertainty by itself. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Name the statistic and sample population. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

79. Ensemble percentile

First principle. A quantile of a specified ensemble distribution. This is a semantic constraint, not merely a preferred term. Use two versions of every short example: a readable sentence for the intended audience and a machine-facing description retaining raw identifiers and scope. Review both against the same source. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Confusing 90th percentile with 90 percent event probability changes assertion. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. The 90th percentile rainfall amount is a threshold, not a forecast that rain occurs with probability 90 percent. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Preserve percentile definition. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

80. Threshold probability

First principle. Share or modeled probability of members exceeding a declared event threshold. This is a semantic constraint, not merely a preferred term. The decisive test is not whether the sentence is elegant. It is whether a colleague who did not inspect the source file could reconstruct the same value class, location, period and method without importing assumptions. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Suppressing threshold makes a probability meaningless. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. Probability of precipitation exceeding 10 mm in 24 hours differs from probability of any precipitation. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. State event threshold, interval and member basis. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

81. Ensemble count

First principle. Number of included, valid ensemble members in a statistic. This is a semantic constraint, not merely a preferred term. Define the origin of authority before choosing the translated noun. Scientific identifiers, value definitions and time support form the invariant core; natural-language ordering and register are flexible. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Hiding missing members can suggest comparability where sample sizes differ. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A percentile from 12 members may be different in reliability from the same nominal percentile from 50. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Report count and exclusions. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

82. Deterministic versus probabilistic products

First principle. Whether output is a single scenario or a distributional summary. This is a semantic constraint, not merely a preferred term. Ask which property is asserted, which quantities are merely metadata and which limitations prevent stronger conclusions. Words such as measured, estimated, averaged, predicted and derived are not substitutes. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Calling every forecast probabilistic implies calibrated likelihoods that may not exist. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A single deterministic run gives a model trajectory, not a probability distribution by itself. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Keep product classification. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

83. Bias correction

First principle. A statistical transformation intended to adjust systematic model differences. This is a semantic constraint, not merely a preferred term. Use two versions of every short example: a readable sentence for the intended audience and a machine-facing description retaining raw identifiers and scope. Review both against the same source. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Calling corrected data direct observations hides a derived product. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A precipitation distribution adjusted to a reference baseline remains a model-derived dataset. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Document training period, method and residual limits. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

84. Downscaling

First principle. A method for deriving finer-resolution fields using dynamics or statistics. This is a semantic constraint, not merely a preferred term. The decisive test is not whether the sentence is elegant. It is whether a colleague who did not inspect the source file could reconstruct the same value class, location, period and method without importing assumptions. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Calling downscaled cells new sensor measurements creates false precision. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A 1 km downscaled map may originate from coarser reanalysis and ancillary predictors. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. State method and source resolution. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

85. Extreme-event attribution

First principle. A structured estimation of how conditions influence event likelihood or intensity. This is a semantic constraint, not merely a preferred term. Define the origin of authority before choosing the translated noun. Scientific identifiers, value definitions and time support form the invariant core; natural-language ordering and register are flexible. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Treating attribution as proof that one event had a single cause misstates uncertainty. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. An estimated risk ratio is conditional on model design and event definition. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Preserve conditional phrasing and uncertainty. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

86. GRIB2 discipline and parameter

First principle. A coded identification system for the type of physical parameter. This is a semantic constraint, not merely a preferred term. Ask which property is asserted, which quantities are merely metadata and which limitations prevent stronger conclusions. Words such as measured, estimated, averaged, predicted and derived are not substitutes. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Translating a parameter label without discipline and category risks a code collision. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A numeric parameter value is not globally unique without its table context. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Keep code-table path with descriptive label. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

87. GRIB2 product definition template

First principle. A structured template describing forecast type and statistical processing. This is a semantic constraint, not merely a preferred term. Use two versions of every short example: a readable sentence for the intended audience and a machine-facing description retaining raw identifiers and scope. Review both against the same source. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Ignoring product definition loses ensemble or time-window semantics. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. Two messages with the same unit and parameter may have different product templates. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Record template and decoded fields. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

88. GRIB2 grid definition

First principle. A coded description of the spatial grid and scanning order. This is a semantic constraint, not merely a preferred term. The decisive test is not whether the sentence is elegant. It is whether a colleague who did not inspect the source file could reconstruct the same value class, location, period and method without importing assumptions. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Translating grid labels while dropping scanning direction can misplace arrays. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. Two grid products with equal point counts can represent different map projections. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Preserve grid definition and scan mode. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

89. GRIB2 local tables

First principle. Codes assigned by a originating centre in a locally scoped table. This is a semantic constraint, not merely a preferred term. Define the origin of authority before choosing the translated noun. Scientific identifiers, value definitions and time support form the invariant core; natural-language ordering and register are flexible. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Treating local code values as universal WMO definitions introduces wrong parameters. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A centre-specific numeric parameter must be resolved against the centre and local table. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Record producer and local-table version. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

90. GRIB2 packing and precision

First principle. Methods that encode numerical arrays with specified scaling and representation. This is a semantic constraint, not merely a preferred term. Ask which property is asserted, which quantities are merely metadata and which limitations prevent stronger conclusions. Words such as measured, estimated, averaged, predicted and derived are not substitutes. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Assuming stored integers are the physical values changes units and range. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. Binary packed fields may need decimal and binary scale factors to recover values. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Decode before translation or arithmetic. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

91. GRIB2 bitmap

First principle. A mask indicating availability or missing grid-cell values. This is a semantic constraint, not merely a preferred term. Use two versions of every short example: a readable sentence for the intended audience and a machine-facing description retaining raw identifiers and scope. Review both against the same source. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Rendering absent cells as zeros invents meteorology. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A missing ocean cell in a land variable is not measured zero moisture. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Preserve bitmap and null semantics. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

92. CF standard_name

First principle. A controlled identifier for the physical quantity of a netCDF variable. This is a semantic constraint, not merely a preferred term. The decisive test is not whether the sentence is elegant. It is whether a colleague who did not inspect the source file could reconstruct the same value class, location, period and method without importing assumptions. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Translating standard_name itself can break CF interoperability. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. Use a localized long_name while leaving a valid standard_name token unchanged. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Separate machine identifiers and prose descriptions. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

93. CF long_name

First principle. A free-text human-readable description of a variable. This is a semantic constraint, not merely a preferred term. Define the origin of authority before choosing the translated noun. Scientific identifiers, value definitions and time support form the invariant core; natural-language ordering and register are flexible. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Treating long_name as an authority over conflicting units and coordinates conceals errors. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A polished label cannot override an inconsistent standard_name or bounds value. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Cross-check all metadata fields. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

94. CF units

First principle. A unit-string attribute interpreted with recognized unit conventions. This is a semantic constraint, not merely a preferred term. Ask which property is asserted, which quantities are merely metadata and which limitations prevent stronger conclusions. Words such as measured, estimated, averaged, predicted and derived are not substitutes. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Replacing native unit strings with nonstandard translated words may break software. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. Keep a machine-recognizable K or Pa in units while describing kelvin or pascal in another language. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Keep units parseable and translations outside the attribute. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

95. CF cell_methods

First principle. A structured attribute identifying aggregation over particular axes. This is a semantic constraint, not merely a preferred term. Use two versions of every short example: a readable sentence for the intended audience and a machine-facing description retaining raw identifiers and scope. Review both against the same source. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Localizing keywords inside the attribute destroys machine semantics. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A method such as time: mean differs from time: sum and cannot be replaced by aesthetic synonyms. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Preserve controlled tokens and localized explanation. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

96. CF time bounds

First principle. A boundary variable giving the start and end of time intervals. This is a semantic constraint, not merely a preferred term. The decisive test is not whether the sentence is elegant. It is whether a colleague who did not inspect the source file could reconstruct the same value class, location, period and method without importing assumptions. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Discarding bounds makes means and sums harder to interpret. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A daily sample indexed at midnight can represent the preceding interval when bounds say so. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Retain boundary variable and verify convention. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

97. CF flag_values and flag_meanings

First principle. Machine-readable associations between values and defined categories. This is a semantic constraint, not merely a preferred term. Define the origin of authority before choosing the translated noun. Scientific identifiers, value definitions and time support form the invariant core; natural-language ordering and register are flexible. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Translating flag_meanings tokens in place may damage downstream parsing. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A new label for missing or suspect should be exposed separately without reassigning numeric flags. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Preserve mappings and provide an external legend. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

98. CF history metadata

First principle. A free-text record of processing steps where available. This is a semantic constraint, not merely a preferred term. Ask which property is asserted, which quantities are merely metadata and which limitations prevent stronger conclusions. Words such as measured, estimated, averaged, predicted and derived are not substitutes. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Deleting preprocessing history hides regridding, merges and conversions. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A derived daily file should identify the source retrieval and transformation choices. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Retain processing trace and append a transparent localization note. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

99. Dataset citation and DOI

First principle. A stable bibliographic reference for a specific data product or publication. This is a semantic constraint, not merely a preferred term. Use two versions of every short example: a readable sentence for the intended audience and a machine-facing description retaining raw identifiers and scope. Review both against the same source. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Changing citation while retaining data makes provenance unrecoverable. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A digital object identifier belongs to the cited source and should not be translated or recomposed. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Preserve exact DOI and access context. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

100. Checksum and integrity

First principle. A digest or other mechanism for confirming file-byte identity. This is a semantic constraint, not merely a preferred term. The decisive test is not whether the sentence is elegant. It is whether a colleague who did not inspect the source file could reconstruct the same value class, location, period and method without importing assumptions. The translator should locate the relevant machine-readable field and an authoritative product explanation before rewriting prose. If one of those is absent, record an unresolved question instead of deciding from an appealing English gloss.

Hidden failure. Describing a transformed file as byte-identical misrepresents integrity checks. A reader may still see a plausible map, table or sentence even though the source assertion has been altered. This makes the error particularly damaging: it tends to survive spell-check, ordinary machine-translation quality metrics and superficial copyediting. A professional review therefore marks the exact missing qualifier, identifier or transformation, then asks whether its omission could alter a scientific conclusion or a practical decision.

Illustrative diagnostic. A translated metadata file may have a new checksum although the original raw file was untouched. To repair such a statement, locate the original variable, identify its physical support, retain the source value and its native unit, and make the consequential distinction explicit in the target language. Do not infer an unrecorded observation, a more exact place, a finer timestamp or a probability solely because those details would make an explanation sound confident.

Reconstruction and transfer test. Keep separate integrity records for source and derivative. The reviewer should write down what evidence would falsify the translated claim, then repeat the check on a different product whose metadata looks superficially similar. A translated title may remain simple for general readers, but a nearby caption, tooltip or methods note must restore information needed to prevent a mistaken comparison. Keep the source snapshot and the decision log so that a later editor can distinguish translation judgement from a source-data revision.

The repeatable method: twenty-eight gates from intake to release

Gate 01 Identify the user and consequential decision

List audience, target language, reading conditions, technical stakes and who can approve a claim. An educational map requires less operational granularity than a warning; neither can distort data. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Gate 02 Freeze the source snapshot

Record the source URL or file, access timestamp, content hash where appropriate, product identifier and dataset edition. Avoid comparing a final translation against a silently changed download. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Gate 03 Establish the authority chain

Distinguish WMO coding authority, CF metadata convention, producer documentation, conversion-tool behaviour and a downstream app’s labels. Resolve conflicts visibly. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Gate 04 Inventory the machine-facing fields

Extract variable, unit, coordinates, bounds, grid mapping, forecast metadata, statistical processing and flags. Mark which are required for science and cannot be localized in place. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Gate 05 Decide the content boundary

Separate immutable data arrays and controlled identifiers from editable titles, captions, prose guidance and reader examples. Localized tooltips should not overwrite standard names. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Gate 06 Reconstruct time before words

Derive reference time, valid time, generation time, lead units, interval bounds, time zone and calendar. Distinguish instant, average, accumulation, extreme and climatology. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Gate 07 Decode GRIB2 product metadata

Use the applicable discipline, category, parameter, product definition, grid definition, generating centre and local-table context. Do not trust one exported abbreviation. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Gate 08 Read CF metadata as a system

Check standard_name, long_name, units, cell_methods, bounds, calendars, formula_terms and grid_mapping together, not as isolated strings. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Gate 09 Confirm measurement class

Label the quantity as observed, analysis estimate, reanalysis, forecast, ensemble statistic, bias-corrected derivative or scenario output as supported. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Gate 10 Confirm native unit

Preserve unit syntax, dimensionality, prefixes, difference-versus-absolute distinctions and any water-equivalent or energy-flux basis. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Gate 11 State spatial support

Mark whether values represent a point, grid cell, area integral, vertical layer, coordinate surface, interpolated field or observational site. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Gate 12 State vertical support

Record pressure levels, height above ground, model levels, depth ranges and reference datums; do not equate model index with altitude. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Gate 13 Resolve vector and sign conventions

For U/V components, flux directions, longitudes and vertical velocities, preserve axis orientation and positive direction before prose. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Gate 14 Resolve aggregation

Identify means, sums, maxima, minima, percentiles, anomalies, climatologies and seasonal calculations with the exact interval or population. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Gate 15 Check ensemble interpretation

Keep control/perturbed member labels, inclusion count, statistic, threshold and calibration qualifiers. Do not translate a quantile into an event probability. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Gate 16 Check missing data

Preserve bitmap, fill values and flags. Keep absent, masked, outside-domain, not applicable and genuine zero separate. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Gate 17 Verify numeric display transformations

Write explicit formulas for Kelvin-to-Celsius, metres-to-millimetres or Pa-to-hPa when legitimately needed. Preserve the original array. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Gate 18 Recheck spatial transformations

If reprojected, interpolated or aggregated, record source grid, target grid, method, area weighting and the relevant conservation test. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Gate 19 Build a bilingual glossary

For each domain term store concept, source term, target form, forbidden near-synonyms and a representative sentence. Keep machine identifiers verbatim. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Gate 20 Draft plain-language explanation

Use readable sentences that communicate the claim without inflating certainty. Put essential qualifiers near the number, not hidden in a distant footnote. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Gate 21 Perform bilingual semantic revision

Have a qualified reviewer compare source and target assertions while inspecting variables and supporting metadata, not only prose. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Gate 22 Perform automated structural checks

Compare numerical tokens, code strings, DOIs, timestamps, units, interval arithmetic, coordinates and output links. Log exceptions rather than silently normalizing. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Gate 23 Test reconstructed meaning

Ask a second reader to infer which quantity, interval, space, provenance and uncertainty a caption refers to; compare their reconstruction with the source. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Gate 24 Run negative-case testing

Intentionally try missing/zero confusion, midnight boundaries, sign flips, code collisions, staggered grids and ensemble-label swaps. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Gate 25 Score translation defects

Define categories for source-fact errors, temporal shifts, spatial shifts, quantification errors, ungrounded certainty, identifier damage and stylistic issues. Weight by harm. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Gate 26 Review accessibility and interface fit

Check that localized axis legends, tooltips, screen-reader text and maps provide understandable units, scope and uncertainty without truncating the crucial qualifier. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Gate 27 Record reviewer decisions

Preserve issue ID, source evidence, rejected wording, approved wording, reviewer, date and whether a data or translation correction is involved. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Gate 28 Release and monitor

Publish only the approved text, verify the rendered page, link the authoritative sources, note version and re-review whenever the dataset or code table changes. Why it matters: a release gate converts an intuitive translation choice into a decision another person can reproduce. Record the evidence rather than an unsupported confidence statement. The gate is complete only when a reviewer can state what was checked, what ambiguity was excluded, what remains unknown and what target-language wording expresses the bounded claim. If the check cannot be performed, retain a visible limitation and route the issue back to the relevant source owner rather than improvise a convenient value.

Failure modes: how apparently correct translations lose scientific evidence

Failure 01: The timestamp is translated but the UTC marker disappears

Mechanism. An audience reads a reference time as a local occurrence time. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Put Z or UTC next to the original timestamp and localize any display conversion in a separate field. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 02: The step is presented as a clock hour

Mechanism. A lead value of 18 becomes 18:00 despite a different reference cycle. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Compute validity from reference and unit before writing the caption. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 03: An interval field becomes an instant

Mechanism. A daily rain total is described as rain measured precisely at midnight. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Display the start and end of accumulation as a range. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 04: An hourly mean becomes a point

Mechanism. A temperature average is compared to an instantaneous station reading. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Keep the time mean and sampling window explicit. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 05: A temporal maximum loses its window

Mechanism. A gust maximum is presented as an exact peak timestamp. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. State the maximum was within a specified interval. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 06: Cumulative precipitation is summed twice

Mechanism. Adjacent outputs share portions of the same accumulation interval. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Inspect step ranges and derive independent increments only when justified. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 07: Accumulation reset is ignored

Mechanism. A new model run is differenced against the last cycle. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Split cycles before calculating differences. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 08: A rate becomes a depth without integration

Mechanism. Intensity is reported as a total despite an unknown duration. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Preserve rate units until a time integral is specified. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 09: Kelvin becomes degrees Kelvin

Mechanism. A unit is reformulated informally and parsed incorrectly. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Use K in data and kelvin in explanatory prose. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 10: A temperature anomaly is treated as an absolute temperature

Mechanism. A plus two K anomaly is shown as two degrees Celsius actual weather. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Keep anomaly and baseline separate from absolute measurements. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 11: Pa and hPa are numerically conflated

Mechanism. A pressure of 101325 Pa is displayed as 101325 hPa. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Use the explicit 100 Pa per hPa relationship. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 12: Surface and sea-level pressure are merged

Mechanism. A mountain cell is compared to MSLP without reduction. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Name the vertical reference and any reduction algorithm. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 13: Pressure level is mistaken for metres

Mechanism. 500 hPa is described as a 500-metre altitude layer. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Keep pressure units and the pressure coordinate. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 14: Model level is translated as height

Mechanism. A level number becomes a fabricated metre value. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Use formula_terms and actual geophysical data to estimate height where appropriate. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 15: Grid cell is described as a weather station

Mechanism. A model result acquires invented site precision. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. State model grid support and comparison limitations. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 16: Longitude is wrapped without coordinate mapping

Mechanism. A point near the date line appears at the wrong map location. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Use reversible zero-to-360 to minus-180-to-180 normalization. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 17: Latitude row order is reversed

Mechanism. North-south positions are flipped while labels remain plausible. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Test index-to-coordinate mapping before plotting. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 18: Grid mapping disappears

Mechanism. Projected x/y numbers are treated as geodetic coordinates. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Preserve projection and coordinate reference description. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 19: Rotated grid angle is ignored

Mechanism. U/V fields are called east/north when relative to model axes. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Check vector orientation and rotation metadata. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 20: Cell bounds are omitted

Mechanism. A time-mean series no longer discloses its temporal intervals. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Carry bounds and cell_methods together. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 21: Cell-area weights are omitted

Mechanism. A global mean is biased by latitudinal grid geometry. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Use appropriate area weighting and document exclusions. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 22: Land fraction becomes a binary coastline

Mechanism. Mixed coastal cells are presented as wholly land or sea. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Keep fractional mask and grid resolution. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 23: Wind components become wind speed

Mechanism. A signed component is converted to an unsigned magnitude by wording. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. State components and derive magnitude only with both components. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 24: Wind from-direction becomes to-direction

Mechanism. The public-facing arrow points into the wind’s origin. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Name convention and verify mapping. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 25: Gust and sustained speed are conflated

Mechanism. An extreme brief peak is reported as mean wind. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Retain peak or average duration. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 26: Water equivalent becomes snow depth

Mechanism. Snowpack water mass is misread as physical depth. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Keep water-equivalent qualifier and density dependence. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 27: Convective and total rain are equated

Mechanism. A model component is overgeneralized to all precipitation. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Explain process components and sum rules. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 28: Runoff depth becomes river discharge

Mechanism. A grid-area water balance acquires an unsupported flow rate. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Keep units and spatial routing assumptions. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 29: Flux sign is flipped

Mechanism. Surface energy budget interpretation reverses. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Check the producer’s sign convention. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 30: A radiative total becomes mean power

Mechanism. Joules per area is plotted as watts per area without duration. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Integrate or divide only with known interval. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 31: Percent cloud cover becomes rain chance

Mechanism. Cloud fraction is wrongly treated as precipitation probability. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Keep separate parameter definitions. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 32: Ensemble control becomes best estimate

Mechanism. One member is falsely identified as privileged by probability. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Name control as configuration rather than ranking. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 33: Ensemble mean becomes a member trajectory

Mechanism. A composite statistic is treated as a realizable sequence. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. State it is a member aggregation. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 34: Spread becomes measurement error

Mechanism. Member dispersion is equated with calibrated observational uncertainty. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Describe statistic and what it does not measure. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 35: Quantile becomes probability

Mechanism. A 90th percentile value is restated as a 90 percent chance of rain. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Explain the quantile threshold and member distribution. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 36: Probability loses threshold

Mechanism. A 60 percent chance of exceeding 10 mm becomes 60 percent chance of any rain. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Keep event definition, window and threshold. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 37: Member count is missing

Mechanism. Comparisons hide changes from 12 to 50 valid members. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Report included and missing members. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 38: Bias-corrected data are called observations

Mechanism. Statistical adjustment is hidden. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Document baseline, method and status as derived. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 39: Downscaled grid is called new measurement

Mechanism. Finer pixels imply unsupported measurement precision. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Retain source grid and technique. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 40: Local GRIB table is treated as universal

Mechanism. A numeric code is assigned the wrong physical variable. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Resolve centre and local-table version. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 41: Product template is lost

Mechanism. Ensemble and accumulation identity vanish in conversion. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Inspect product definition and translated representation. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 42: Bitmap mask is converted to zero

Mechanism. Missing cells become false dry or cold areas. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Preserve missingness separately. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 43: CF identifiers are translated in place

Mechanism. Software can no longer parse controlled vocabulary tokens. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Keep standard_name and cell_methods machine-stable. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 44: Data and metadata versions diverge

Mechanism. A new caption describes an old array under new assumptions. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Bind publication to source release and processing history. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 45: Chart legend drops uncertainty

Mechanism. Readers treat a probability range as a certainty statement. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Show threshold, ensemble and uncertainty in accessible language. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 46: Source citation is replaced by a generic homepage

Mechanism. Readers cannot retrieve the precise dataset edition. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Link dataset record and DOI when present. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 47: Generated illustrative examples are presented as real events

Mechanism. Fictional measurements become cited weather facts. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Clearly label demonstrations as hypothetical. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Failure 48: Machine translation is accepted without domain review

Mechanism. A fluent sentence hides unit or temporal mistakes. This error may persist because the translated phrase is grammatically natural and the numeric value still looks reasonable. The analyst should not accept fluency as evidence. Locate the raw data and any time, area, unit, processing, ensemble or source fields that would make this sentence false under an alternative reading. A common trigger is a tool that exports only a short variable label while hiding the source metadata.

Correction and audit. Use bilingual and scientific checks before release. The revised target-language explanation must carry enough information to distinguish the safe interpretation from the misleading one. Add a short diagnostic test to the QA record: reconstruct the value’s referent from the final caption, compare against the source, and ensure a reader cannot reasonably infer the original error. If an interpretation remains product-dependent, identify that dependency rather than asserting an unverified universal rule.

Worked diagnostic library: construct, challenge and repair technical claims

The cases below are fictional exercises. They deliberately use common-looking values and short source claims rather than reproducing an agency’s current forecast. Their role is to show how a rigorous translator builds a defensible statement and a testable error record. A real translation should always inspect the actual producing centre’s definitions and file metadata, which take precedence over any simplified illustration.

Diagnostic 001: Analysis field in research methods paragraph

Source scenario (hypothetical). A team prepares a peer-reviewed methods section that contains analysis field and a short original note: “Hypothetical temperature analysis of 289 K at 00 UTC is not itself an individual thermometer measurement”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Replacing estimated with directly observed erases modelling and analysis context. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the dataset edition and computational derivation. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Write analysis estimate, state time, and keep its provenance separate from station observations. A usable target-language note should explain that an estimated atmospheric state at an analysis time, often informed by data assimilation and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable an independent scientist to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 002: Operational forecast in public climate dashboard

Source scenario (hypothetical). A team prepares a general-reader chart caption that contains operational forecast and a short original note: “A 12-hour forecast initialized at 00 UTC can be valid at 12 UTC although it was produced earlier”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Calling a forecast an observation falsely upgrades evidence. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the time window and value class. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Pair initialization with valid time and retain forecast designation. A usable target-language note should explain that a model prediction tied to an initialization and a lead time and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a reader who has not used GRIB2 to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 003: Reanalysis product in multilingual data dictionary

Source scenario (hypothetical). A team prepares a software variable glossary that contains reanalysis product and a short original note: “An ERA5 hourly value estimates a past state over a grid rather than reproducing one station record”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Calling reanalysis raw historical weather observations hides assimilation and uncertainty. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to canonical identifiers and units. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Retain product name, release and reanalysis classification. A usable target-language note should explain that a retrospective, model-assisted reconstruction using observations and a consistent production system and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a developer validating the schema to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 004: Hindcast in operational forecast note

Source scenario (hypothetical). A team prepares a time-sensitive but non-warning summary that contains hindcast and a short original note: “A seasonal forecast model run retrospectively for old years may not assimilate observations like a reanalysis”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Confusing hindcast and reanalysis loses how observations were used. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to reference and validity times. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Document simulation design and whether a run was initialized with historical inputs. A usable target-language note should explain that a retrospective model prediction or simulation used to evaluate behaviour over historical periods and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a forecaster verifying a bulletin to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 005: Nowcast in education explainer

Source scenario (hypothetical). A team prepares an introductory science lesson that contains nowcast and a short original note: “An extrapolated precipitation radar field may represent the next short interval rather than an instrument observation”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Treating nowcast as a verified current observation overclaims certainty. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the key distinction without fabricated precision. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Preserve product label and latency or forecast horizon. A usable target-language note should explain that a very short-range estimate or forecast with product-specific methodology and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a learner solving a data-interpretation task to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 006: Climate projection in data publication readme

Source scenario (hypothetical). A team prepares an archive-level methodological description that contains climate projection and a short original note: “An ensemble under one emissions pathway is not a prediction of one afternoon’s weather”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Calling a scenario-dependent projection a weather forecast suggests different predictive certainty. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to source and transformation provenance. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Preserve scenario, model, period and conditional language. A usable target-language note should explain that a conditional future model result based on scenarios and assumptions and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a future data reuser to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 007: Climatological normal in research methods paragraph

Source scenario (hypothetical). A team prepares a peer-reviewed methods section that contains climatological normal and a short original note: “A monthly temperature normal over a 30-year baseline is a reference statistic, not a physiological ideal”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Using normal to imply natural or desirable changes the claim. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the dataset edition and computational derivation. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. State statistic and baseline dates. A usable target-language note should explain that a statistic calculated for a specified multi-year reference interval and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable an independent scientist to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 008: Climate anomaly in public climate dashboard

Source scenario (hypothetical). A team prepares a general-reader chart caption that contains climate anomaly and a short original note: “An anomaly of plus 1.2 K requires the reader to know the reference period and statistic”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Printing the departure without its baseline makes comparison ambiguous. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the time window and value class. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Translate anomaly as difference from the named baseline. A usable target-language note should explain that a departure relative to a stated reference climatology or baseline and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a reader who has not used GRIB2 to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 009: Dataset edition in multilingual data dictionary

Source scenario (hypothetical). A team prepares a software variable glossary that contains dataset edition and a short original note: “Two releases may share variable names while differing in bias correction or timeliness”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Mixing values or descriptions from two releases can manufacture apparent trends. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to canonical identifiers and units. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Preserve version, release date and retrieval date. A usable target-language note should explain that the identity of a delivered product and its particular published version and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a developer validating the schema to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 010: Provisional release in operational forecast note

Source scenario (hypothetical). A team prepares a time-sensitive but non-warning summary that contains provisional release and a short original note: “A recent reanalysis product labelled provisional must not be described as a frozen historical record”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Suppressing provisional status portrays revision-prone data as final. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to reference and validity times. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Keep provisional qualifier and the expected revision pathway without inventing dates. A usable target-language note should explain that a near-real-time release with a documented path to later replacement or adjustment and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a forecaster verifying a bulletin to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 011: Analysis reference time in education explainer

Source scenario (hypothetical). A team prepares an introductory science lesson that contains analysis reference time and a short original note: “A run reference 2026-04-08T00Z plus a 24-hour step has a different valid time”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Treating reference time as the time represented by any forecast step misplaces events. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the key distinction without fabricated precision. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Store both clocks and their relation. A usable target-language note should explain that the timestamp associated with an analysis or initialization and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a learner solving a data-interpretation task to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 012: Forecast validity time in data publication readme

Source scenario (hypothetical). A team prepares an archive-level methodological description that contains forecast validity time and a short original note: “An accumulated field valid at 18 UTC may summarize hours 12 through 18”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Describing valid time as the start of any accumulation creates a shifted timeline. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to source and transformation provenance. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Display validity and interval bounds separately. A usable target-language note should explain that the time at which an instantaneous forecast applies or an interval field is indexed and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a future data reuser to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 013: Forecast lead in research methods paragraph

Source scenario (hypothetical). A team prepares a peer-reviewed methods section that contains forecast lead and a short original note: “The same lead duration may be encoded in minutes or hours”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Assuming a bare 12 always means twelve hours ignores code-table time-unit choices. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the dataset edition and computational derivation. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Decode unit and numerical lead before prose translation. A usable target-language note should explain that elapsed forecast time from the reference cycle, with units defined by the product and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable an independent scientist to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 014: Production time in public climate dashboard

Source scenario (hypothetical). A team prepares a general-reader chart caption that contains production time and a short original note: “A forecast copied the next day retains its original initialization and validity”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Confusing file creation with phenomenon time makes stale products appear current. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the time window and value class. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Keep generation timestamp as distribution metadata. A usable target-language note should explain that when a file was generated, assembled, or delivered and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a reader who has not used GRIB2 to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 015: Time-zone indication in multilingual data dictionary

Source scenario (hypothetical). A team prepares a software variable glossary that contains time-zone indication and a short original note: “An ISO timestamp ending in Z is UTC, not a clock reading in every reader’s locale”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Adding an unqualified local-time label shifts the temporal referent. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to canonical identifiers and units. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Preserve ISO timestamp and display locale conversion as an additional view. A usable target-language note should explain that a declared convention for interpreting timestamps and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a developer validating the schema to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 016: Calendar attribute in operational forecast note

Source scenario (hypothetical). A team prepares a time-sensitive but non-warning summary that contains calendar attribute and a short original note: “A model month with thirty days in a 360_day calendar is not necessarily a real calendar month”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Applying Gregorian arithmetic to a 360_day model calendar produces false dates. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to reference and validity times. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Keep calendar attribute and describe model-time conventions. A usable target-language note should explain that the calendar model used by a netCDF time coordinate and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a forecaster verifying a bulletin to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 017: Time coordinate units in education explainer

Source scenario (hypothetical). A team prepares an introductory science lesson that contains time coordinate units and a short original note: “A time coordinate 18 with units hours since a UTC date encodes an offset, not hour-of-day by itself”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Reading hours since a base date as an absolute timestamp produces nonsense. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the key distinction without fabricated precision. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Decode units and epoch together. A usable target-language note should explain that the numeric epoch and duration units used for time values and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a learner solving a data-interpretation task to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 018: Instantaneous sample in data publication readme

Source scenario (hypothetical). A team prepares an archive-level methodological description that contains instantaneous sample and a short original note: “Surface pressure at 09 UTC is not automatically an average between 08 and 09 UTC”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Calling a sample an hourly mean claims an averaging operation that never occurred. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to source and transformation provenance. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Use at time unless interval processing is documented. A usable target-language note should explain that a value representative of a particular time or instant and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a future data reuser to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 019: Temporal mean in research methods paragraph

Source scenario (hypothetical). A team prepares a peer-reviewed methods section that contains temporal mean and a short original note: “A daily mean from 24 hourly fields depends on which 24 fields and time basis were selected”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Presenting mean temperature as a point observation conceals its time support. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the dataset edition and computational derivation. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Preserve averaging method and interval bounds. A usable target-language note should explain that a statistic over a declared interval with beginning and end and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable an independent scientist to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 020: Temporal sum in public climate dashboard

Source scenario (hypothetical). A team prepares a general-reader chart caption that contains temporal sum and a short original note: “A six-hour water equivalent of 0.012 m represents 12 mm over that interval, not 12 mm per hour”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Confusing total depth with rate can inflate totals by hours or seconds. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the time window and value class. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Report total, duration and interval endpoints. A usable target-language note should explain that a quantity accumulated or summed across a time interval and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a reader who has not used GRIB2 to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 021: Temporal maximum in multilingual data dictionary

Source scenario (hypothetical). A team prepares a software variable glossary that contains temporal maximum and a short original note: “A maximum wind gust over three hours cannot be assigned to the window endpoint without source evidence”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Writing maximum as instantaneous value misrepresents duration. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to canonical identifiers and units. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Preserve window and method. A usable target-language note should explain that the greatest reported value within a stated window and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a developer validating the schema to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 022: Temporal minimum in operational forecast note

Source scenario (hypothetical). A team prepares a time-sensitive but non-warning summary that contains temporal minimum and a short original note: “A six-hour minimum temperature indexed by 06 UTC may have occurred earlier”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Inferring the exact occurrence instant from window validity adds unsupported precision. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to reference and validity times. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Keep window and avoid inventing event time. A usable target-language note should explain that the lowest reported value within a stated window and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a forecaster verifying a bulletin to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 023: Rolling window in education explainer

Source scenario (hypothetical). A team prepares an introductory science lesson that contains rolling window and a short original note: “Consecutive twelve-hour rainfall totals overlapping by eleven hours cannot be added as independent totals”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Summing overlapping rolling windows double-counts the same underlying hours. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the key distinction without fabricated precision. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Mark window length, movement and overlap. A usable target-language note should explain that a statistic calculated for a window that moves between successive outputs and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a learner solving a data-interpretation task to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 024: Accumulation reset in data publication readme

Source scenario (hypothetical). A team prepares an archive-level methodological description that contains accumulation reset and a short original note: “The zero-hour value and next cycle’s accumulation are not necessarily continuous”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Differencing across a reset may fabricate a negative rainfall amount. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to source and transformation provenance. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Check cycle boundaries before deriving increments. A usable target-language note should explain that the point at which a run or product resets a cumulative quantity and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a future data reuser to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 025: Accumulation beginning in research methods paragraph

Source scenario (hypothetical). A team prepares a peer-reviewed methods section that contains accumulation beginning and a short original note: “A 12 to 18 hour forecast sum begins at hour 12, not at forecast hour zero”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Assuming beginning equals initialization loses partial-window products. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the dataset edition and computational derivation. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Record beginning lead and ending lead. A usable target-language note should explain that the first time in the temporal support of an accumulated field and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable an independent scientist to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 026: Accumulation end in public climate dashboard

Source scenario (hypothetical). A team prepares a general-reader chart caption that contains accumulation end and a short original note: “A forecast labelled valid 18 UTC may cover the preceding six hours”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Treating the endpoint as a single-instant observation changes the value class. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the time window and value class. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Explain end time and duration. A usable target-language note should explain that the endpoint that indexes a completed accumulation and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a reader who has not used GRIB2 to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 027: Forecast step range in multilingual data dictionary

Source scenario (hypothetical). A team prepares a software variable glossary that contains forecast step range and a short original note: “The intervals 0 to 18 and 12 to 18 both end at hour 18 but differ materially”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Translating only the latter number hides whether the statistic was accumulated from zero. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to canonical identifiers and units. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Retain both step boundaries. A usable target-language note should explain that two lead-time endpoints used to identify the supported interval and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a developer validating the schema to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 028: Accumulation versus rate in operational forecast note

Source scenario (hypothetical). A team prepares a time-sensitive but non-warning summary that contains accumulation versus rate and a short original note: “A precipitation rate of 0.001 kg m−2 s−1 must not be described as one millimetre without a duration”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Adding rates as if they were totals ignores integration. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to reference and validity times. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. State units, integration window and conversion assumptions. A usable target-language note should explain that the distinction between total over a window and change per unit time and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a forecaster verifying a bulletin to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 029: Kelvin temperature in education explainer

Source scenario (hypothetical). A team prepares an introductory science lesson that contains kelvin temperature and a short original note: “A temperature of 293.15 K is 20 degrees Celsius, whereas a difference of 1 K equals a difference of 1 Celsius degree”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Appending degrees or treating a change like an absolute temperature causes errors. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the key distinction without fabricated precision. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Preserve absolute versus difference semantics. A usable target-language note should explain that an absolute thermodynamic temperature with symbol K and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a learner solving a data-interpretation task to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 030: Celsius display in data publication readme

Source scenario (hypothetical). A team prepares an archive-level methodological description that contains celsius display and a short original note: “A map legend converted from Kelvin to Celsius must document subtracting 273.15 from absolute temperature”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Changing values without labelling the converted units breaks reproducibility. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to source and transformation provenance. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Keep native array and make display transform explicit. A usable target-language note should explain that a derived presentation of temperature from a compatible absolute source and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a future data reuser to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 031: Surface pressure in research methods paragraph

Source scenario (hypothetical). A team prepares a peer-reviewed methods section that contains surface pressure and a short original note: “A high-elevation grid cell can have lower surface pressure without a storm”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Calling surface pressure sea-level pressure disregards elevation and reduction method. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the dataset edition and computational derivation. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Retain surface reference and Pa or hPa unit. A usable target-language note should explain that atmospheric pressure at the model surface or terrain height and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable an independent scientist to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 032: Mean sea-level pressure in public climate dashboard

Source scenario (hypothetical). A team prepares a general-reader chart caption that contains mean sea-level pressure and a short original note: “An MSLP contour differs semantically from a station’s actual surface pressure”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Treating it as measured at ground level creates a false instrument comparison. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the time window and value class. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Retain reduction concept. A usable target-language note should explain that pressure adjusted to a reference at mean sea level using a defined method and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a reader who has not used GRIB2 to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 033: Geopotential in multilingual data dictionary

Source scenario (hypothetical). A team prepares a software variable glossary that contains geopotential and a short original note: “A 500 hPa geopotential field has unit m2 s−2 in some products and may be displayed as geopotential height after division by gravity”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Equating geopotential with geometric altitude without transformation can misstate vertical positions. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to canonical identifiers and units. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Document source parameter and any converted product. A usable target-language note should explain that potential energy per unit mass associated with gravity in meteorology and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a developer validating the schema to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 034: Pressure level in operational forecast note

Source scenario (hypothetical). A team prepares a time-sensitive but non-warning summary that contains pressure level and a short original note: “Temperature at 500 hPa represents the specified pressure surface which varies in height”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Interpreting 500 hPa as 500 metres is a categorical unit error. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to reference and validity times. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Preserve hPa label and variable’s vertical coordinate. A usable target-language note should explain that an isobaric vertical surface identified by pressure rather than geometric height and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a forecaster verifying a bulletin to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 035: Hybrid model level in education explainer

Source scenario (hypothetical). A team prepares an introductory science lesson that contains hybrid model level and a short original note: “The geometric height of level 40 depends on pressure and surface terms, not the ordinal 40 alone”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Translating a model level number into metres invents altitude. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the key distinction without fabricated precision. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Retain formula_terms and model-level identity. A usable target-language note should explain that a terrain-following vertical coordinate defined by formula terms and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a learner solving a data-interpretation task to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 036: Height above ground in data publication readme

Source scenario (hypothetical). A team prepares an archive-level methodological description that contains height above ground and a short original note: “A 10 m wind is not a wind measured at arbitrary rooftop height”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Omitting the reference turns 2 m air temperature into an undefined vertical observation. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to source and transformation provenance. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Preserve height and datum. A usable target-language note should explain that an explicit vertical displacement relative to ground or a specified surface and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a future data reuser to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 037: Grid-cell mean in research methods paragraph

Source scenario (hypothetical). A team prepares a peer-reviewed methods section that contains grid-cell mean and a short original note: “A coarse grid covers variable terrain and cannot promise a thermometer reading at one address”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Writing at a named city can imply exact local measurement. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the dataset edition and computational derivation. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Use over the grid cell and note spatial resolution. A usable target-language note should explain that a quantity representative of a spatial cell or support, not a specific point sensor and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable an independent scientist to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 038: Grid-point value in public climate dashboard

Source scenario (hypothetical). A team prepares a general-reader chart caption that contains grid-point value and a short original note: “A point interpolation and cell mean have different supports even at similar displayed coordinates”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Describing one point as a regional average can change representativeness. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the time window and value class. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Record interpolation and grid definition. A usable target-language note should explain that a quantity evaluated at a discrete grid coordinate with model support and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a reader who has not used GRIB2 to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 039: Latitude ordering in multilingual data dictionary

Source scenario (hypothetical). A team prepares a software variable glossary that contains latitude ordering and a short original note: “A stored latitude sequence 90, 89, 88 is descending and must match array indices”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Reversing row order while keeping latitude labels unchanged flips a map vertically. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to canonical identifiers and units. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Validate coordinate monotonicity and bounds. A usable target-language note should explain that whether array rows run north to south or south to north and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a developer validating the schema to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 040: Longitude convention in operational forecast note

Source scenario (hypothetical). A team prepares a time-sensitive but non-warning summary that contains longitude convention and a short original note: “Longitude 350 degrees east is equivalent to minus 10 degrees, not a site ten degrees east”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Treating an equivalent wrapped longitude as a different physical position duplicates points. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to reference and validity times. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Normalize only for display with reversible mapping. A usable target-language note should explain that whether longitudes are represented as zero to 360 or minus 180 to 180 and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a forecaster verifying a bulletin to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 041: Rotated grid in education explainer

Source scenario (hypothetical). A team prepares an introductory science lesson that contains rotated grid and a short original note: “A rotated-pole grid needs its grid mapping to locate a cell correctly”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Labelling rotated coordinates directly as Earth latitude and longitude shifts locations. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the key distinction without fabricated precision. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Preserve rotated coordinate definition. A usable target-language note should explain that a coordinate grid whose axes differ from ordinary geodetic latitude and longitude and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a learner solving a data-interpretation task to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 042: Projected grid in data publication readme

Source scenario (hypothetical). A team prepares an archive-level methodological description that contains projected grid and a short original note: “A Lambert conformal grid requires its standard parallels and projection origin for reproducibility”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Relabelling projected x/y as latitude/longitude destroys geographic meaning. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to source and transformation provenance. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Carry projection metadata and coordinates. A usable target-language note should explain that a coordinate system defined by map projection and its parameters and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a future data reuser to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 043: Grid mapping in research methods paragraph

Source scenario (hypothetical). A team prepares a peer-reviewed methods section that contains grid mapping and a short original note: “An image may look visually plausible while the underlying transform has vanished”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Deleting the mapping can make values appear in a different region. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the dataset edition and computational derivation. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Retain grid_mapping and a separately localized description. A usable target-language note should explain that cF metadata connecting native coordinates to a coordinate reference system and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable an independent scientist to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 044: Cell bounds in public climate dashboard

Source scenario (hypothetical). A team prepares a general-reader chart caption that contains cell bounds and a short original note: “CF bounds can locate temporal or spatial support more precisely than coordinate labels alone”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Assuming a point coordinate always means the centre of a rectangular cell introduces unsupported geometry. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the time window and value class. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Preserve bounds and do not invent centroids. A usable target-language note should explain that coordinates delineating a cell’s interval or spatial area and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a reader who has not used GRIB2 to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 045: Cell area in multilingual data dictionary

Source scenario (hypothetical). A team prepares a software variable glossary that contains cell area and a short original note: “One degree of longitude spans different physical distances at different latitudes”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Averaging unequal-area cells without weights can bias a global statistic. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to canonical identifiers and units. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Retain cell_measures or appropriate area weights. A usable target-language note should explain that area used for spatial weighting or integration and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a developer validating the schema to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 046: Land-sea mask in operational forecast note

Source scenario (hypothetical). A team prepares a time-sensitive but non-warning summary that contains land-sea mask and a short original note: “A coastal cell with land fraction 0.4 contains both surface types, not a perfectly mapped shoreline”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Treating partial land fraction as an absolute coastline invents resolution. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to reference and validity times. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Explain fractional or categorical mask and grid support. A usable target-language note should explain that a field identifying land/ocean fraction or classification in a cell and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a forecaster verifying a bulletin to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 047: Oroography in education explainer

Source scenario (hypothetical). A team prepares an introductory science lesson that contains oroography and a short original note: “A grid elevation is an aggregate or model-representative height”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Describing model surface altitude as precise local topography falsely sharpens the data. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the key distinction without fabricated precision. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Label terrain representation and source resolution. A usable target-language note should explain that representation of surface elevation in a model grid and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a learner solving a data-interpretation task to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 048: Interpolation method in data publication readme

Source scenario (hypothetical). A team prepares an archive-level methodological description that contains interpolation method and a short original note: “Nearest neighbour and bilinear interpolation can produce different extreme values”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Calling regridded numbers native measurements conceals a derived calculation. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to source and transformation provenance. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Record source grid, target grid and method. A usable target-language note should explain that a transformation estimating values onto new coordinates and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a future data reuser to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 049: Regridding conservativeness in research methods paragraph

Source scenario (hypothetical). A team prepares a peer-reviewed methods section that contains regridding conservativeness and a short original note: “Conservative remapping is a property to verify, not a guarantee from a translated label”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Applying a generic interpolator to area totals may change integrated quantities. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the dataset edition and computational derivation. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Keep method and check integrated totals. A usable target-language note should explain that whether a spatial transform conserves an extensive quantity under stated assumptions and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable an independent scientist to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 050: Wind U component in public climate dashboard

Source scenario (hypothetical). A team prepares a general-reader chart caption that contains wind u component and a short original note: “U equals minus 5 m/s describes a westward component in an east-positive coordinate system”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Rendering U as wind speed ignores vector direction and possible grid rotation. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the time window and value class. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. State component orientation and sign. A usable target-language note should explain that the eastward or grid-relative component of wind velocity and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a reader who has not used GRIB2 to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 051: Wind V component in multilingual data dictionary

Source scenario (hypothetical). A team prepares a software variable glossary that contains wind v component and a short original note: “A V value of plus 4 m/s is a component, not proof of a four metre per second total wind”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Replacing V with a plain speed loses vector orientation. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to canonical identifiers and units. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Keep coordinate basis and component label. A usable target-language note should explain that the northward or grid-relative wind component and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a developer validating the schema to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 052: Wind speed magnitude in operational forecast note

Source scenario (hypothetical). A team prepares a time-sensitive but non-warning summary that contains wind speed magnitude and a short original note: “A speed of 10 m/s can arise from different directions and cannot reveal its own components”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Taking arithmetic means of magnitudes as vector-mean wind changes statistics. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to reference and validity times. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Preserve magnitude and aggregation method. A usable target-language note should explain that the magnitude calculated from vector components or supplied by a model and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a forecaster verifying a bulletin to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 053: Wind direction convention in education explainer

Source scenario (hypothetical). A team prepares an introductory science lesson that contains wind direction convention and a short original note: “A meteorological northerly wind comes from the north, not travels toward it”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Flipping from-direction to to-direction changes meaning by 180 degrees. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the key distinction without fabricated precision. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Declare convention and angle reference. A usable target-language note should explain that the convention for direction from which or toward which wind travels and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a learner solving a data-interpretation task to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 054: Wind gust in data publication readme

Source scenario (hypothetical). A team prepares an archive-level methodological description that contains wind gust and a short original note: “A maximum gust of 22 m/s cannot be read as a ten-minute average without evidence”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Treating a gust as mean wind speed exaggerates sustained conditions. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to source and transformation provenance. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Preserve peak definition and window. A usable target-language note should explain that a short-duration peak under a model or observational definition and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a future data reuser to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 055: Dew point in research methods paragraph

Source scenario (hypothetical). A team prepares a peer-reviewed methods section that contains dew point and a short original note: “Air temperature 28 C and dew point 21 C are different fields even when a chart shows temperatures”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Mistaking dew point for air temperature changes humidity interpretation. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the dataset edition and computational derivation. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Retain physical definition. A usable target-language note should explain that temperature at which air reaches saturation for a given moisture state and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable an independent scientist to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 056: Specific humidity in public climate dashboard

Source scenario (hypothetical). A team prepares a general-reader chart caption that contains specific humidity and a short original note: “0.010 kg/kg of moist air does not directly equal a percent relative humidity”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Confusing specific humidity and mixing ratio changes denominator. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the time window and value class. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Preserve exact property and unit. A usable target-language note should explain that mass of water vapour per mass of moist air and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a reader who has not used GRIB2 to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 057: Relative humidity in multilingual data dictionary

Source scenario (hypothetical). A team prepares a software variable glossary that contains relative humidity and a short original note: “80 percent RH is not 80 percent water by mass”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Treating percent relative humidity as mass percentage creates a false material balance. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to canonical identifiers and units. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Keep saturation-based definition and percent unit. A usable target-language note should explain that vapour pressure relative to saturation under a defined temperature basis and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a developer validating the schema to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 058: Total cloud cover in operational forecast note

Source scenario (hypothetical). A team prepares a time-sensitive but non-warning summary that contains total cloud cover and a short original note: “A cover fraction of 0.75 means 75 percent under the product’s definition, not a 75 percent precipitation chance”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Treating fractional cover as cloud thickness or rain probability changes variables. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to reference and validity times. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Label coverage and support. A usable target-language note should explain that fraction of sky or grid area represented by cloud cover under a product definition and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a forecaster verifying a bulletin to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 059: Convective precipitation in education explainer

Source scenario (hypothetical). A team prepares an introductory science lesson that contains convective precipitation and a short original note: “A model separates convective and large-scale contributions by parameterization, which differs by system”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Translating the category as all thunderstorm rainfall may overstate physical attribution. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the key distinction without fabricated precision. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Retain parameter distinction. A usable target-language note should explain that precipitation parameter attributed to a model’s convective scheme and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a learner solving a data-interpretation task to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 060: Large-scale precipitation in data publication readme

Source scenario (hypothetical). A team prepares an archive-level methodological description that contains large-scale precipitation and a short original note: “A labelled component is defined by scheme, not a public claim about affected area”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Equating large-scale with wide geographic extent is too literal. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to source and transformation provenance. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Explain model process classification. A usable target-language note should explain that precipitation attributed to resolved cloud or large-scale processes in the model and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a future data reuser to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 061: Total precipitation in research methods paragraph

Source scenario (hypothetical). A team prepares a peer-reviewed methods section that contains total precipitation and a short original note: “A product total of 0.008 m water equivalent represents eight millimetres for its declared window”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Replacing it with rain amount discards snow or other included frozen precipitation. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the dataset edition and computational derivation. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Keep inclusions, units and accumulation bounds. A usable target-language note should explain that water-equivalent precipitation amount over a defined interval and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable an independent scientist to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 062: Snowfall water equivalent in public climate dashboard

Source scenario (hypothetical). A team prepares a general-reader chart caption that contains snowfall water equivalent and a short original note: “Ten millimetres water equivalent is not necessarily ten millimetres snow depth”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Equating it directly with snow depth ignores density. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the time window and value class. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Identify water-equivalent basis. A usable target-language note should explain that liquid-water equivalent corresponding to a snow quantity and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a reader who has not used GRIB2 to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 063: Snow depth in multilingual data dictionary

Source scenario (hypothetical). A team prepares a software variable glossary that contains snow depth and a short original note: “A snowpack 0.30 m deep can have much smaller water equivalent depending on density”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Misreading depth as water equivalent breaks storage estimates. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to canonical identifiers and units. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Keep physical depth and unit. A usable target-language note should explain that the modelled or measured vertical depth of snow cover and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a developer validating the schema to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 064: Surface runoff in operational forecast note

Source scenario (hypothetical). A team prepares a time-sensitive but non-warning summary that contains surface runoff and a short original note: “A runoff depth per grid area does not equal cubic metres per second at a gauge”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Calling it observed river discharge changes location and unit. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to reference and validity times. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Report depth and spatial basis. A usable target-language note should explain that water that leaves the land surface as flow under a model’s water balance and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a forecaster verifying a bulletin to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 065: Subsurface runoff in education explainer

Source scenario (hypothetical). A team prepares an introductory science lesson that contains subsurface runoff and a short original note: “A stored drainage depth may contribute to routing without equalling an aquifer recharge observation”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Calling all subsurface runoff groundwater recharge ignores routing and definitions. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the key distinction without fabricated precision. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Preserve process and timescale. A usable target-language note should explain that modelled below-surface drainage component over a given time and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a learner solving a data-interpretation task to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 066: Soil moisture layer in data publication readme

Source scenario (hypothetical). A team prepares an archive-level methodological description that contains soil moisture layer and a short original note: “A value in m3/m3 is dimensionless fraction rather than millimetres of water”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Conflating volumetric water content with depth-integrated water storage changes units. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to source and transformation provenance. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Preserve layer bounds and water-content basis. A usable target-language note should explain that water content stored in a specified soil depth or model layer and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a future data reuser to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 067: Shortwave downward radiation in research methods paragraph

Source scenario (hypothetical). A team prepares a peer-reviewed methods section that contains shortwave downward radiation and a short original note: “An accumulated surface energy of 3.6 MJ m−2 over one hour corresponds to a mean flux of 1000 W m−2 under exact constant-interval assumptions”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Confusing J m−2 totals with W m−2 rates changes magnitude and duration. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the dataset edition and computational derivation. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Report direction, unit and interval. A usable target-language note should explain that solar-origin radiation incident on a surface, expressed as flux or accumulated energy and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable an independent scientist to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 068: Longwave radiation in public climate dashboard

Source scenario (hypothetical). A team prepares a general-reader chart caption that contains longwave radiation and a short original note: “Downwelling and upwelling longwave terms are physically distinct”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Removing upward or downward can reverse an energy-budget interpretation. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the time window and value class. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Keep radiation band and direction. A usable target-language note should explain that thermal infrared radiative flux or accumulated energy with directional specification and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a reader who has not used GRIB2 to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 069: Sensible heat flux in multilingual data dictionary

Source scenario (hypothetical). A team prepares a software variable glossary that contains sensible heat flux and a short original note: “A negative flux in one product may represent upward energy under a stated convention, while another uses opposite sign”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Dropping sign convention can reverse surface heating interpretation. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to canonical identifiers and units. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Check positive-direction rule. A usable target-language note should explain that turbulent transfer of heat without phase-change latent energy and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a developer validating the schema to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 070: Latent heat flux in operational forecast note

Source scenario (hypothetical). A team prepares a time-sensitive but non-warning summary that contains latent heat flux and a short original note: “An energy flux needs duration and latent-heat assumptions to estimate water mass exchange”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Treating it as a measured evaporation depth without latent heat conversion is invalid. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to reference and validity times. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Preserve energy form and sign. A usable target-language note should explain that energy associated with moisture phase change and turbulent exchange and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a forecaster verifying a bulletin to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 071: CAPE in education explainer

Source scenario (hypothetical). A team prepares an introductory science lesson that contains cape and a short original note: “A CAPE value in J/kg indicates parcel buoyancy energy, not a direct event probability”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Calling CAPE a probability of storms is a category error. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the key distinction without fabricated precision. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Preserve diagnostic nature and units. A usable target-language note should explain that convective available potential energy of an atmospheric parcel under specified assumptions and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a learner solving a data-interpretation task to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 072: Boundary-layer height in data publication readme

Source scenario (hypothetical). A team prepares an archive-level methodological description that contains boundary-layer height and a short original note: “A mixed-layer depth can differ from cloud-base or inversion height”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Reporting it as the altitude of every cloud base invents a relationship. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to source and transformation provenance. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Retain diagnostic definition. A usable target-language note should explain that a model-derived estimate of atmospheric boundary-layer depth and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a future data reuser to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 073: Visibility in research methods paragraph

Source scenario (hypothetical). A team prepares a peer-reviewed methods section that contains visibility and a short original note: “A model field may use extinction estimates rather than direct human observation”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Interpreting visual range as proof of clear air under all conditions overclaims. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the dataset edition and computational derivation. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Specify method and unit. A usable target-language note should explain that a visibility diagnostic or observation with model-dependent definition and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable an independent scientist to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 074: Ozone mixing ratio in public climate dashboard

Source scenario (hypothetical). A team prepares a general-reader chart caption that contains ozone mixing ratio and a short original note: “Ozone at a pressure level differs from surface concentration used in exposure assessment”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Treating all ozone fields as surface air quality changes altitude and relevance. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the time window and value class. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Preserve vertical level and composition basis. A usable target-language note should explain that an atmospheric composition quantity with a specific concentration or mixing-ratio unit and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a reader who has not used GRIB2 to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 075: Ensemble control in multilingual data dictionary

Source scenario (hypothetical). A team prepares a software variable glossary that contains ensemble control and a short original note: “Control is a member label, not automatically a probability maximum”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Calling the control member the most likely outcome misstates ensemble interpretation. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to canonical identifiers and units. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Keep member identity and generation description. A usable target-language note should explain that a designated member produced with the system’s control configuration and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a developer validating the schema to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 076: Perturbed member in operational forecast note

Source scenario (hypothetical). A team prepares a time-sensitive but non-warning summary that contains perturbed member and a short original note: “Member 17 is an identifier, not seventeenth-best forecast”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Treating member number as ranking invents an order. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to reference and validity times. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Preserve member number without probability claims. A usable target-language note should explain that an ensemble realization with an intentional perturbation in initial conditions or model representation and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a forecaster verifying a bulletin to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 077: Ensemble mean in education explainer

Source scenario (hypothetical). A team prepares an introductory science lesson that contains ensemble mean and a short original note: “An ensemble-mean rainfall amount is not necessarily realized by any one member”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Treating the mean as an actual member’s trajectory may create impossible combinations. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the key distinction without fabricated precision. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. State aggregation set and statistic. A usable target-language note should explain that a statistic across selected members at a specified time and variable and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a learner solving a data-interpretation task to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 078: Ensemble spread in data publication readme

Source scenario (hypothetical). A team prepares an archive-level methodological description that contains ensemble spread and a short original note: “A standard deviation of member temperatures is not an observational measurement uncertainty by itself”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Translating spread as error or confidence interval without method overstates inference. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to source and transformation provenance. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Name the statistic and sample population. A usable target-language note should explain that a dispersion statistic whose definition may be standard deviation or another measure and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a future data reuser to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 079: Ensemble percentile in research methods paragraph

Source scenario (hypothetical). A team prepares a peer-reviewed methods section that contains ensemble percentile and a short original note: “The 90th percentile rainfall amount is a threshold, not a forecast that rain occurs with probability 90 percent”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Confusing 90th percentile with 90 percent event probability changes assertion. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the dataset edition and computational derivation. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Preserve percentile definition. A usable target-language note should explain that a quantile of a specified ensemble distribution and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable an independent scientist to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 080: Threshold probability in public climate dashboard

Source scenario (hypothetical). A team prepares a general-reader chart caption that contains threshold probability and a short original note: “Probability of precipitation exceeding 10 mm in 24 hours differs from probability of any precipitation”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Suppressing threshold makes a probability meaningless. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the time window and value class. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. State event threshold, interval and member basis. A usable target-language note should explain that share or modeled probability of members exceeding a declared event threshold and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a reader who has not used GRIB2 to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 081: Ensemble count in multilingual data dictionary

Source scenario (hypothetical). A team prepares a software variable glossary that contains ensemble count and a short original note: “A percentile from 12 members may be different in reliability from the same nominal percentile from 50”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Hiding missing members can suggest comparability where sample sizes differ. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to canonical identifiers and units. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Report count and exclusions. A usable target-language note should explain that number of included, valid ensemble members in a statistic and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a developer validating the schema to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 082: Deterministic versus probabilistic products in operational forecast note

Source scenario (hypothetical). A team prepares a time-sensitive but non-warning summary that contains deterministic versus probabilistic products and a short original note: “A single deterministic run gives a model trajectory, not a probability distribution by itself”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Calling every forecast probabilistic implies calibrated likelihoods that may not exist. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to reference and validity times. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Keep product classification. A usable target-language note should explain that whether output is a single scenario or a distributional summary and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a forecaster verifying a bulletin to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 083: Bias correction in education explainer

Source scenario (hypothetical). A team prepares an introductory science lesson that contains bias correction and a short original note: “A precipitation distribution adjusted to a reference baseline remains a model-derived dataset”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Calling corrected data direct observations hides a derived product. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the key distinction without fabricated precision. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Document training period, method and residual limits. A usable target-language note should explain that a statistical transformation intended to adjust systematic model differences and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a learner solving a data-interpretation task to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 084: Downscaling in data publication readme

Source scenario (hypothetical). A team prepares an archive-level methodological description that contains downscaling and a short original note: “A 1 km downscaled map may originate from coarser reanalysis and ancillary predictors”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Calling downscaled cells new sensor measurements creates false precision. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to source and transformation provenance. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. State method and source resolution. A usable target-language note should explain that a method for deriving finer-resolution fields using dynamics or statistics and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a future data reuser to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 085: Extreme-event attribution in research methods paragraph

Source scenario (hypothetical). A team prepares a peer-reviewed methods section that contains extreme-event attribution and a short original note: “An estimated risk ratio is conditional on model design and event definition”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Treating attribution as proof that one event had a single cause misstates uncertainty. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the dataset edition and computational derivation. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Preserve conditional phrasing and uncertainty. A usable target-language note should explain that a structured estimation of how conditions influence event likelihood or intensity and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable an independent scientist to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 086: GRIB2 discipline and parameter in public climate dashboard

Source scenario (hypothetical). A team prepares a general-reader chart caption that contains grib2 discipline and parameter and a short original note: “A numeric parameter value is not globally unique without its table context”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Translating a parameter label without discipline and category risks a code collision. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the time window and value class. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Keep code-table path with descriptive label. A usable target-language note should explain that a coded identification system for the type of physical parameter and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a reader who has not used GRIB2 to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 087: GRIB2 product definition template in multilingual data dictionary

Source scenario (hypothetical). A team prepares a software variable glossary that contains grib2 product definition template and a short original note: “Two messages with the same unit and parameter may have different product templates”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Ignoring product definition loses ensemble or time-window semantics. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to canonical identifiers and units. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Record template and decoded fields. A usable target-language note should explain that a structured template describing forecast type and statistical processing and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a developer validating the schema to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 088: GRIB2 grid definition in operational forecast note

Source scenario (hypothetical). A team prepares a time-sensitive but non-warning summary that contains grib2 grid definition and a short original note: “Two grid products with equal point counts can represent different map projections”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Translating grid labels while dropping scanning direction can misplace arrays. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to reference and validity times. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Preserve grid definition and scan mode. A usable target-language note should explain that a coded description of the spatial grid and scanning order and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a forecaster verifying a bulletin to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 089: GRIB2 local tables in education explainer

Source scenario (hypothetical). A team prepares an introductory science lesson that contains grib2 local tables and a short original note: “A centre-specific numeric parameter must be resolved against the centre and local table”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Treating local code values as universal WMO definitions introduces wrong parameters. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the key distinction without fabricated precision. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Record producer and local-table version. A usable target-language note should explain that codes assigned by a originating centre in a locally scoped table and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a learner solving a data-interpretation task to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 090: GRIB2 packing and precision in data publication readme

Source scenario (hypothetical). A team prepares an archive-level methodological description that contains grib2 packing and precision and a short original note: “Binary packed fields may need decimal and binary scale factors to recover values”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Assuming stored integers are the physical values changes units and range. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to source and transformation provenance. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Decode before translation or arithmetic. A usable target-language note should explain that methods that encode numerical arrays with specified scaling and representation and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a future data reuser to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 091: GRIB2 bitmap in research methods paragraph

Source scenario (hypothetical). A team prepares a peer-reviewed methods section that contains grib2 bitmap and a short original note: “A missing ocean cell in a land variable is not measured zero moisture”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Rendering absent cells as zeros invents meteorology. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the dataset edition and computational derivation. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Preserve bitmap and null semantics. A usable target-language note should explain that a mask indicating availability or missing grid-cell values and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable an independent scientist to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 092: CF standard_name in public climate dashboard

Source scenario (hypothetical). A team prepares a general-reader chart caption that contains cf standard_name and a short original note: “Use a localized long_name while leaving a valid standard_name token unchanged”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Translating standard_name itself can break CF interoperability. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the time window and value class. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Separate machine identifiers and prose descriptions. A usable target-language note should explain that a controlled identifier for the physical quantity of a netCDF variable and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a reader who has not used GRIB2 to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 093: CF long_name in multilingual data dictionary

Source scenario (hypothetical). A team prepares a software variable glossary that contains cf long_name and a short original note: “A polished label cannot override an inconsistent standard_name or bounds value”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Treating long_name as an authority over conflicting units and coordinates conceals errors. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to canonical identifiers and units. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Cross-check all metadata fields. A usable target-language note should explain that a free-text human-readable description of a variable and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a developer validating the schema to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 094: CF units in operational forecast note

Source scenario (hypothetical). A team prepares a time-sensitive but non-warning summary that contains cf units and a short original note: “Keep a machine-recognizable K or Pa in units while describing kelvin or pascal in another language”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Replacing native unit strings with nonstandard translated words may break software. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to reference and validity times. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Keep units parseable and translations outside the attribute. A usable target-language note should explain that a unit-string attribute interpreted with recognized unit conventions and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a forecaster verifying a bulletin to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 095: CF cell_methods in education explainer

Source scenario (hypothetical). A team prepares an introductory science lesson that contains cf cell_methods and a short original note: “A method such as time: mean differs from time: sum and cannot be replaced by aesthetic synonyms”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Localizing keywords inside the attribute destroys machine semantics. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the key distinction without fabricated precision. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Preserve controlled tokens and localized explanation. A usable target-language note should explain that a structured attribute identifying aggregation over particular axes and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a learner solving a data-interpretation task to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 096: CF time bounds in data publication readme

Source scenario (hypothetical). A team prepares an archive-level methodological description that contains cf time bounds and a short original note: “A daily sample indexed at midnight can represent the preceding interval when bounds say so”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Discarding bounds makes means and sums harder to interpret. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to source and transformation provenance. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Retain boundary variable and verify convention. A usable target-language note should explain that a boundary variable giving the start and end of time intervals and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a future data reuser to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 097: CF flag_values and flag_meanings in research methods paragraph

Source scenario (hypothetical). A team prepares a peer-reviewed methods section that contains cf flag_values and flag_meanings and a short original note: “A new label for missing or suspect should be exposed separately without reassigning numeric flags”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Translating flag_meanings tokens in place may damage downstream parsing. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the dataset edition and computational derivation. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Preserve mappings and provide an external legend. A usable target-language note should explain that machine-readable associations between values and defined categories and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable an independent scientist to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 098: CF history metadata in public climate dashboard

Source scenario (hypothetical). A team prepares a general-reader chart caption that contains cf history metadata and a short original note: “A derived daily file should identify the source retrieval and transformation choices”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Deleting preprocessing history hides regridding, merges and conversions. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to the time window and value class. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Retain processing trace and append a transparent localization note. A usable target-language note should explain that a free-text record of processing steps where available and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a reader who has not used GRIB2 to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 099: Dataset citation and DOI in multilingual data dictionary

Source scenario (hypothetical). A team prepares a software variable glossary that contains dataset citation and doi and a short original note: “A digital object identifier belongs to the cited source and should not be translated or recomposed”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Changing citation while retaining data makes provenance unrecoverable. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to canonical identifiers and units. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Preserve exact DOI and access context. A usable target-language note should explain that a stable bibliographic reference for a specific data product or publication and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a developer validating the schema to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Diagnostic 100: Checksum and integrity in operational forecast note

Source scenario (hypothetical). A team prepares a time-sensitive but non-warning summary that contains checksum and integrity and a short original note: “A translated metadata file may have a new checksum although the original raw file was untouched”. This note is not a live observation or an official warning. The record also has a source dataset identifier, a UTC timestamp and a machine-readable unit; the translator initially receives only the prose. The dangerous shortcut is to translate the phrase without requesting the properties that distinguish this specific record from a superficially similar one.

Naïve interpretation and diagnosis. Describing a transformed file as byte-identical misrepresents integrity checks. A literal sentence can remain fluent while performing that substitution. Before rewriting it, the reviewer must identify which part of the sentence is a measurement, which part names a processing choice, what the time and spatial support are and whether any inference exceeds the data. To make the decision reproducible, write a yes/no check keyed to reference and validity times. If the relevant field is not supplied, the result is “needs source metadata”, not “translator may guess”.

Evidence-preserving rendering. Keep separate integrity records for source and derivative. A usable target-language note should explain that a digest or other mechanism for confirming file-byte identity and retain the source file’s exact identifiers in a separate code-facing track. A caption may be shorter than a paper’s methods section, but it cannot omit the qualifier that changes the numerical interpretation. The translated version should enable a forecaster verifying a bulletin to tell what the claim refers to and what it does not establish.

Independent test and extension. Give a second reviewer the original schema and the final prose, with all editorial notes hidden. Ask them to identify the supported property, original units, temporal and spatial referents, product status, and any qualifying uncertainty. Then alter only one source field—such as the processing interval, model cycle, ensemble statistic or unit—and ask whether the translated statement must change. If the wording survives a source alteration that changes the scientific meaning, the translation was underspecified. If it changes despite an irrelevant alteration, it may be overfitted to one file.

Twenty-four transfer workshops: apply the system beyond one dataset

Workshop 01: Reanalysis temperature against a station log

Begin with a hypothetical source package containing two versions of the same figure: an internal scientific chart and a public explanatory chart. The question is not whether both can share attractive wording, but whether their audiences need different layers of explanation without a different scientific claim. For the reanalysis temperature against a station log case, first inspect the original record and identify which facts would survive a translation into a third language. Establish whether the data are an observation, model estimate, prediction, derived product or statistic; then list all units, times, spaces and version identifiers needed to reconstruct the figure.

Construct a comparison table with four columns: original machine field, source-language explanation, target-language explanation and independent QA evidence. Give special attention to analysis field because replacing estimated with directly observed erases modelling and analysis context. Cross-check climate anomaly as a second variable; Translate anomaly as difference from the named baseline. The exercise succeeds only when the technical distinctions are preserved while the target-language wording becomes easier to understand. An interface label can be brief if an accessible tooltip supplies the excluded details.

For the adversarial test, introduce one deliberate change in time support, coordinate reference, unit or ensemble meaning while leaving the visible numeric token unchanged. Ask another reader to determine whether the old caption is still true. A complete review records the mistake that was detected, the source evidence that resolved it, the approved wording, and the limit of what the chart can establish. Finally, apply the same protocol to a product from another centre. The aim is transfer of the method, not memorisation of one provider’s parameter names.

Workshop 02: Forecast precipitation against a rain gauge

Begin with a hypothetical source package containing two versions of the same figure: an internal scientific chart and a public explanatory chart. The question is not whether both can share attractive wording, but whether their audiences need different layers of explanation without a different scientific claim. For the forecast precipitation against a rain gauge case, first inspect the original record and identify which facts would survive a translation into a third language. Establish whether the data are an observation, model estimate, prediction, derived product or statistic; then list all units, times, spaces and version identifiers needed to reconstruct the figure.

Construct a comparison table with four columns: original machine field, source-language explanation, target-language explanation and independent QA evidence. Give special attention to hindcast because confusing hindcast and reanalysis loses how observations were used. Cross-check analysis reference time as a second variable; Store both clocks and their relation. The exercise succeeds only when the technical distinctions are preserved while the target-language wording becomes easier to understand. An interface label can be brief if an accessible tooltip supplies the excluded details.

For the adversarial test, introduce one deliberate change in time support, coordinate reference, unit or ensemble meaning while leaving the visible numeric token unchanged. Ask another reader to determine whether the old caption is still true. A complete review records the mistake that was detected, the source evidence that resolved it, the approved wording, and the limit of what the chart can establish. Finally, apply the same protocol to a product from another centre. The aim is transfer of the method, not memorisation of one provider’s parameter names.

Workshop 03: GRIB2 binary message to netCDF export

Begin with a hypothetical source package containing two versions of the same figure: an internal scientific chart and a public explanatory chart. The question is not whether both can share attractive wording, but whether their audiences need different layers of explanation without a different scientific claim. For the grib2 binary message to netcdf export case, first inspect the original record and identify which facts would survive a translation into a third language. Establish whether the data are an observation, model estimate, prediction, derived product or statistic; then list all units, times, spaces and version identifiers needed to reconstruct the figure.

Construct a comparison table with four columns: original machine field, source-language explanation, target-language explanation and independent QA evidence. Give special attention to climatological normal because using normal to imply natural or desirable changes the claim. Cross-check production time as a second variable; Keep generation timestamp as distribution metadata. The exercise succeeds only when the technical distinctions are preserved while the target-language wording becomes easier to understand. An interface label can be brief if an accessible tooltip supplies the excluded details.

For the adversarial test, introduce one deliberate change in time support, coordinate reference, unit or ensemble meaning while leaving the visible numeric token unchanged. Ask another reader to determine whether the old caption is still true. A complete review records the mistake that was detected, the source evidence that resolved it, the approved wording, and the limit of what the chart can establish. Finally, apply the same protocol to a product from another centre. The aim is transfer of the method, not memorisation of one provider’s parameter names.

Workshop 04: Daily mean to monthly climatology

Begin with a hypothetical source package containing two versions of the same figure: an internal scientific chart and a public explanatory chart. The question is not whether both can share attractive wording, but whether their audiences need different layers of explanation without a different scientific claim. For the daily mean to monthly climatology case, first inspect the original record and identify which facts would survive a translation into a third language. Establish whether the data are an observation, model estimate, prediction, derived product or statistic; then list all units, times, spaces and version identifiers needed to reconstruct the figure.

Construct a comparison table with four columns: original machine field, source-language explanation, target-language explanation and independent QA evidence. Give special attention to provisional release because suppressing provisional status portrays revision-prone data as final. Cross-check time coordinate units as a second variable; Decode units and epoch together. The exercise succeeds only when the technical distinctions are preserved while the target-language wording becomes easier to understand. An interface label can be brief if an accessible tooltip supplies the excluded details.

For the adversarial test, introduce one deliberate change in time support, coordinate reference, unit or ensemble meaning while leaving the visible numeric token unchanged. Ask another reader to determine whether the old caption is still true. A complete review records the mistake that was detected, the source evidence that resolved it, the approved wording, and the limit of what the chart can establish. Finally, apply the same protocol to a product from another centre. The aim is transfer of the method, not memorisation of one provider’s parameter names.

Workshop 05: Ensemble member to ensemble mean

Begin with a hypothetical source package containing two versions of the same figure: an internal scientific chart and a public explanatory chart. The question is not whether both can share attractive wording, but whether their audiences need different layers of explanation without a different scientific claim. For the ensemble member to ensemble mean case, first inspect the original record and identify which facts would survive a translation into a third language. Establish whether the data are an observation, model estimate, prediction, derived product or statistic; then list all units, times, spaces and version identifiers needed to reconstruct the figure.

Construct a comparison table with four columns: original machine field, source-language explanation, target-language explanation and independent QA evidence. Give special attention to forecast lead because assuming a bare 12 always means twelve hours ignores code-table time-unit choices. Cross-check temporal sum as a second variable; Report total, duration and interval endpoints. The exercise succeeds only when the technical distinctions are preserved while the target-language wording becomes easier to understand. An interface label can be brief if an accessible tooltip supplies the excluded details.

For the adversarial test, introduce one deliberate change in time support, coordinate reference, unit or ensemble meaning while leaving the visible numeric token unchanged. Ask another reader to determine whether the old caption is still true. A complete review records the mistake that was detected, the source evidence that resolved it, the approved wording, and the limit of what the chart can establish. Finally, apply the same protocol to a product from another centre. The aim is transfer of the method, not memorisation of one provider’s parameter names.

Workshop 06: Surface pressure versus sea-level reduction

Begin with a hypothetical source package containing two versions of the same figure: an internal scientific chart and a public explanatory chart. The question is not whether both can share attractive wording, but whether their audiences need different layers of explanation without a different scientific claim. For the surface pressure versus sea-level reduction case, first inspect the original record and identify which facts would survive a translation into a third language. Establish whether the data are an observation, model estimate, prediction, derived product or statistic; then list all units, times, spaces and version identifiers needed to reconstruct the figure.

Construct a comparison table with four columns: original machine field, source-language explanation, target-language explanation and independent QA evidence. Give special attention to calendar attribute because applying Gregorian arithmetic to a 360_day model calendar produces false dates. Cross-check rolling window as a second variable; Mark window length, movement and overlap. The exercise succeeds only when the technical distinctions are preserved while the target-language wording becomes easier to understand. An interface label can be brief if an accessible tooltip supplies the excluded details.

For the adversarial test, introduce one deliberate change in time support, coordinate reference, unit or ensemble meaning while leaving the visible numeric token unchanged. Ask another reader to determine whether the old caption is still true. A complete review records the mistake that was detected, the source evidence that resolved it, the approved wording, and the limit of what the chart can establish. Finally, apply the same protocol to a product from another centre. The aim is transfer of the method, not memorisation of one provider’s parameter names.

Workshop 07: Wind U/V to speed and direction

Begin with a hypothetical source package containing two versions of the same figure: an internal scientific chart and a public explanatory chart. The question is not whether both can share attractive wording, but whether their audiences need different layers of explanation without a different scientific claim. For the wind u/v to speed and direction case, first inspect the original record and identify which facts would survive a translation into a third language. Establish whether the data are an observation, model estimate, prediction, derived product or statistic; then list all units, times, spaces and version identifiers needed to reconstruct the figure.

Construct a comparison table with four columns: original machine field, source-language explanation, target-language explanation and independent QA evidence. Give special attention to temporal mean because presenting mean temperature as a point observation conceals its time support. Cross-check accumulation end as a second variable; Explain end time and duration. The exercise succeeds only when the technical distinctions are preserved while the target-language wording becomes easier to understand. An interface label can be brief if an accessible tooltip supplies the excluded details.

For the adversarial test, introduce one deliberate change in time support, coordinate reference, unit or ensemble meaning while leaving the visible numeric token unchanged. Ask another reader to determine whether the old caption is still true. A complete review records the mistake that was detected, the source evidence that resolved it, the approved wording, and the limit of what the chart can establish. Finally, apply the same protocol to a product from another centre. The aim is transfer of the method, not memorisation of one provider’s parameter names.

Workshop 08: Latitude-longitude to projected grid

Begin with a hypothetical source package containing two versions of the same figure: an internal scientific chart and a public explanatory chart. The question is not whether both can share attractive wording, but whether their audiences need different layers of explanation without a different scientific claim. For the latitude-longitude to projected grid case, first inspect the original record and identify which facts would survive a translation into a third language. Establish whether the data are an observation, model estimate, prediction, derived product or statistic; then list all units, times, spaces and version identifiers needed to reconstruct the figure.

Construct a comparison table with four columns: original machine field, source-language explanation, target-language explanation and independent QA evidence. Give special attention to temporal minimum because inferring the exact occurrence instant from window validity adds unsupported precision. Cross-check kelvin temperature as a second variable; Preserve absolute versus difference semantics. The exercise succeeds only when the technical distinctions are preserved while the target-language wording becomes easier to understand. An interface label can be brief if an accessible tooltip supplies the excluded details.

For the adversarial test, introduce one deliberate change in time support, coordinate reference, unit or ensemble meaning while leaving the visible numeric token unchanged. Ask another reader to determine whether the old caption is still true. A complete review records the mistake that was detected, the source evidence that resolved it, the approved wording, and the limit of what the chart can establish. Finally, apply the same protocol to a product from another centre. The aim is transfer of the method, not memorisation of one provider’s parameter names.

Workshop 09: Pressure-level map to altitude language

Begin with a hypothetical source package containing two versions of the same figure: an internal scientific chart and a public explanatory chart. The question is not whether both can share attractive wording, but whether their audiences need different layers of explanation without a different scientific claim. For the pressure-level map to altitude language case, first inspect the original record and identify which facts would survive a translation into a third language. Establish whether the data are an observation, model estimate, prediction, derived product or statistic; then list all units, times, spaces and version identifiers needed to reconstruct the figure.

Construct a comparison table with four columns: original machine field, source-language explanation, target-language explanation and independent QA evidence. Give special attention to accumulation beginning because assuming beginning equals initialization loses partial-window products. Cross-check mean sea-level pressure as a second variable; Retain reduction concept. The exercise succeeds only when the technical distinctions are preserved while the target-language wording becomes easier to understand. An interface label can be brief if an accessible tooltip supplies the excluded details.

For the adversarial test, introduce one deliberate change in time support, coordinate reference, unit or ensemble meaning while leaving the visible numeric token unchanged. Ask another reader to determine whether the old caption is still true. A complete review records the mistake that was detected, the source evidence that resolved it, the approved wording, and the limit of what the chart can establish. Finally, apply the same protocol to a product from another centre. The aim is transfer of the method, not memorisation of one provider’s parameter names.

Workshop 10: Water-equivalent snowfall to snow depth

Begin with a hypothetical source package containing two versions of the same figure: an internal scientific chart and a public explanatory chart. The question is not whether both can share attractive wording, but whether their audiences need different layers of explanation without a different scientific claim. For the water-equivalent snowfall to snow depth case, first inspect the original record and identify which facts would survive a translation into a third language. Establish whether the data are an observation, model estimate, prediction, derived product or statistic; then list all units, times, spaces and version identifiers needed to reconstruct the figure.

Construct a comparison table with four columns: original machine field, source-language explanation, target-language explanation and independent QA evidence. Give special attention to accumulation versus rate because adding rates as if they were totals ignores integration. Cross-check hybrid model level as a second variable; Retain formula_terms and model-level identity. The exercise succeeds only when the technical distinctions are preserved while the target-language wording becomes easier to understand. An interface label can be brief if an accessible tooltip supplies the excluded details.

For the adversarial test, introduce one deliberate change in time support, coordinate reference, unit or ensemble meaning while leaving the visible numeric token unchanged. Ask another reader to determine whether the old caption is still true. A complete review records the mistake that was detected, the source evidence that resolved it, the approved wording, and the limit of what the chart can establish. Finally, apply the same protocol to a product from another centre. The aim is transfer of the method, not memorisation of one provider’s parameter names.

Workshop 11: Radiation energy totals to flux language

Begin with a hypothetical source package containing two versions of the same figure: an internal scientific chart and a public explanatory chart. The question is not whether both can share attractive wording, but whether their audiences need different layers of explanation without a different scientific claim. For the radiation energy totals to flux language case, first inspect the original record and identify which facts would survive a translation into a third language. Establish whether the data are an observation, model estimate, prediction, derived product or statistic; then list all units, times, spaces and version identifiers needed to reconstruct the figure.

Construct a comparison table with four columns: original machine field, source-language explanation, target-language explanation and independent QA evidence. Give special attention to surface pressure because calling surface pressure sea-level pressure disregards elevation and reduction method. Cross-check grid-point value as a second variable; Record interpolation and grid definition. The exercise succeeds only when the technical distinctions are preserved while the target-language wording becomes easier to understand. An interface label can be brief if an accessible tooltip supplies the excluded details.

For the adversarial test, introduce one deliberate change in time support, coordinate reference, unit or ensemble meaning while leaving the visible numeric token unchanged. Ask another reader to determine whether the old caption is still true. A complete review records the mistake that was detected, the source evidence that resolved it, the approved wording, and the limit of what the chart can establish. Finally, apply the same protocol to a product from another centre. The aim is transfer of the method, not memorisation of one provider’s parameter names.

Workshop 12: Runoff depth to hydrological discharge

Begin with a hypothetical source package containing two versions of the same figure: an internal scientific chart and a public explanatory chart. The question is not whether both can share attractive wording, but whether their audiences need different layers of explanation without a different scientific claim. For the runoff depth to hydrological discharge case, first inspect the original record and identify which facts would survive a translation into a third language. Establish whether the data are an observation, model estimate, prediction, derived product or statistic; then list all units, times, spaces and version identifiers needed to reconstruct the figure.

Construct a comparison table with four columns: original machine field, source-language explanation, target-language explanation and independent QA evidence. Give special attention to pressure level because interpreting 500 hPa as 500 metres is a categorical unit error. Cross-check rotated grid as a second variable; Preserve rotated coordinate definition. The exercise succeeds only when the technical distinctions are preserved while the target-language wording becomes easier to understand. An interface label can be brief if an accessible tooltip supplies the excluded details.

For the adversarial test, introduce one deliberate change in time support, coordinate reference, unit or ensemble meaning while leaving the visible numeric token unchanged. Ask another reader to determine whether the old caption is still true. A complete review records the mistake that was detected, the source evidence that resolved it, the approved wording, and the limit of what the chart can establish. Finally, apply the same protocol to a product from another centre. The aim is transfer of the method, not memorisation of one provider’s parameter names.

Workshop 13: Cloud fraction versus rain probability

Begin with a hypothetical source package containing two versions of the same figure: an internal scientific chart and a public explanatory chart. The question is not whether both can share attractive wording, but whether their audiences need different layers of explanation without a different scientific claim. For the cloud fraction versus rain probability case, first inspect the original record and identify which facts would survive a translation into a third language. Establish whether the data are an observation, model estimate, prediction, derived product or statistic; then list all units, times, spaces and version identifiers needed to reconstruct the figure.

Construct a comparison table with four columns: original machine field, source-language explanation, target-language explanation and independent QA evidence. Give special attention to grid-cell mean because writing at a named city can imply exact local measurement. Cross-check cell bounds as a second variable; Preserve bounds and do not invent centroids. The exercise succeeds only when the technical distinctions are preserved while the target-language wording becomes easier to understand. An interface label can be brief if an accessible tooltip supplies the excluded details.

For the adversarial test, introduce one deliberate change in time support, coordinate reference, unit or ensemble meaning while leaving the visible numeric token unchanged. Ask another reader to determine whether the old caption is still true. A complete review records the mistake that was detected, the source evidence that resolved it, the approved wording, and the limit of what the chart can establish. Finally, apply the same protocol to a product from another centre. The aim is transfer of the method, not memorisation of one provider’s parameter names.

Workshop 14: Forecast probability versus percentiles

Begin with a hypothetical source package containing two versions of the same figure: an internal scientific chart and a public explanatory chart. The question is not whether both can share attractive wording, but whether their audiences need different layers of explanation without a different scientific claim. For the forecast probability versus percentiles case, first inspect the original record and identify which facts would survive a translation into a third language. Establish whether the data are an observation, model estimate, prediction, derived product or statistic; then list all units, times, spaces and version identifiers needed to reconstruct the figure.

Construct a comparison table with four columns: original machine field, source-language explanation, target-language explanation and independent QA evidence. Give special attention to longitude convention because treating an equivalent wrapped longitude as a different physical position duplicates points. Cross-check oroography as a second variable; Label terrain representation and source resolution. The exercise succeeds only when the technical distinctions are preserved while the target-language wording becomes easier to understand. An interface label can be brief if an accessible tooltip supplies the excluded details.

For the adversarial test, introduce one deliberate change in time support, coordinate reference, unit or ensemble meaning while leaving the visible numeric token unchanged. Ask another reader to determine whether the old caption is still true. A complete review records the mistake that was detected, the source evidence that resolved it, the approved wording, and the limit of what the chart can establish. Finally, apply the same protocol to a product from another centre. The aim is transfer of the method, not memorisation of one provider’s parameter names.

Workshop 15: Anomaly maps with different baselines

Begin with a hypothetical source package containing two versions of the same figure: an internal scientific chart and a public explanatory chart. The question is not whether both can share attractive wording, but whether their audiences need different layers of explanation without a different scientific claim. For the anomaly maps with different baselines case, first inspect the original record and identify which facts would survive a translation into a third language. Establish whether the data are an observation, model estimate, prediction, derived product or statistic; then list all units, times, spaces and version identifiers needed to reconstruct the figure.

Construct a comparison table with four columns: original machine field, source-language explanation, target-language explanation and independent QA evidence. Give special attention to grid mapping because deleting the mapping can make values appear in a different region. Cross-check wind u component as a second variable; State component orientation and sign. The exercise succeeds only when the technical distinctions are preserved while the target-language wording becomes easier to understand. An interface label can be brief if an accessible tooltip supplies the excluded details.

For the adversarial test, introduce one deliberate change in time support, coordinate reference, unit or ensemble meaning while leaving the visible numeric token unchanged. Ask another reader to determine whether the old caption is still true. A complete review records the mistake that was detected, the source evidence that resolved it, the approved wording, and the limit of what the chart can establish. Finally, apply the same protocol to a product from another centre. The aim is transfer of the method, not memorisation of one provider’s parameter names.

Workshop 16: Near-real-time provisional to final release

Begin with a hypothetical source package containing two versions of the same figure: an internal scientific chart and a public explanatory chart. The question is not whether both can share attractive wording, but whether their audiences need different layers of explanation without a different scientific claim. For the near-real-time provisional to final release case, first inspect the original record and identify which facts would survive a translation into a third language. Establish whether the data are an observation, model estimate, prediction, derived product or statistic; then list all units, times, spaces and version identifiers needed to reconstruct the figure.

Construct a comparison table with four columns: original machine field, source-language explanation, target-language explanation and independent QA evidence. Give special attention to land-sea mask because treating partial land fraction as an absolute coastline invents resolution. Cross-check wind direction convention as a second variable; Declare convention and angle reference. The exercise succeeds only when the technical distinctions are preserved while the target-language wording becomes easier to understand. An interface label can be brief if an accessible tooltip supplies the excluded details.

For the adversarial test, introduce one deliberate change in time support, coordinate reference, unit or ensemble meaning while leaving the visible numeric token unchanged. Ask another reader to determine whether the old caption is still true. A complete review records the mistake that was detected, the source evidence that resolved it, the approved wording, and the limit of what the chart can establish. Finally, apply the same protocol to a product from another centre. The aim is transfer of the method, not memorisation of one provider’s parameter names.

Workshop 17: Masked coastal cells and land fraction

Begin with a hypothetical source package containing two versions of the same figure: an internal scientific chart and a public explanatory chart. The question is not whether both can share attractive wording, but whether their audiences need different layers of explanation without a different scientific claim. For the masked coastal cells and land fraction case, first inspect the original record and identify which facts would survive a translation into a third language. Establish whether the data are an observation, model estimate, prediction, derived product or statistic; then list all units, times, spaces and version identifiers needed to reconstruct the figure.

Construct a comparison table with four columns: original machine field, source-language explanation, target-language explanation and independent QA evidence. Give special attention to regridding conservativeness because applying a generic interpolator to area totals may change integrated quantities. Cross-check specific humidity as a second variable; Preserve exact property and unit. The exercise succeeds only when the technical distinctions are preserved while the target-language wording becomes easier to understand. An interface label can be brief if an accessible tooltip supplies the excluded details.

For the adversarial test, introduce one deliberate change in time support, coordinate reference, unit or ensemble meaning while leaving the visible numeric token unchanged. Ask another reader to determine whether the old caption is still true. A complete review records the mistake that was detected, the source evidence that resolved it, the approved wording, and the limit of what the chart can establish. Finally, apply the same protocol to a product from another centre. The aim is transfer of the method, not memorisation of one provider’s parameter names.

Workshop 18: Grid-point value versus cell area average

Begin with a hypothetical source package containing two versions of the same figure: an internal scientific chart and a public explanatory chart. The question is not whether both can share attractive wording, but whether their audiences need different layers of explanation without a different scientific claim. For the grid-point value versus cell area average case, first inspect the original record and identify which facts would survive a translation into a third language. Establish whether the data are an observation, model estimate, prediction, derived product or statistic; then list all units, times, spaces and version identifiers needed to reconstruct the figure.

Construct a comparison table with four columns: original machine field, source-language explanation, target-language explanation and independent QA evidence. Give special attention to wind speed magnitude because taking arithmetic means of magnitudes as vector-mean wind changes statistics. Cross-check convective precipitation as a second variable; Retain parameter distinction. The exercise succeeds only when the technical distinctions are preserved while the target-language wording becomes easier to understand. An interface label can be brief if an accessible tooltip supplies the excluded details.

For the adversarial test, introduce one deliberate change in time support, coordinate reference, unit or ensemble meaning while leaving the visible numeric token unchanged. Ask another reader to determine whether the old caption is still true. A complete review records the mistake that was detected, the source evidence that resolved it, the approved wording, and the limit of what the chart can establish. Finally, apply the same protocol to a product from another centre. The aim is transfer of the method, not memorisation of one provider’s parameter names.

Workshop 19: Local GRIB code table migration

Begin with a hypothetical source package containing two versions of the same figure: an internal scientific chart and a public explanatory chart. The question is not whether both can share attractive wording, but whether their audiences need different layers of explanation without a different scientific claim. For the local grib code table migration case, first inspect the original record and identify which facts would survive a translation into a third language. Establish whether the data are an observation, model estimate, prediction, derived product or statistic; then list all units, times, spaces and version identifiers needed to reconstruct the figure.

Construct a comparison table with four columns: original machine field, source-language explanation, target-language explanation and independent QA evidence. Give special attention to dew point because mistaking dew point for air temperature changes humidity interpretation. Cross-check snowfall water equivalent as a second variable; Identify water-equivalent basis. The exercise succeeds only when the technical distinctions are preserved while the target-language wording becomes easier to understand. An interface label can be brief if an accessible tooltip supplies the excluded details.

For the adversarial test, introduce one deliberate change in time support, coordinate reference, unit or ensemble meaning while leaving the visible numeric token unchanged. Ask another reader to determine whether the old caption is still true. A complete review records the mistake that was detected, the source evidence that resolved it, the approved wording, and the limit of what the chart can establish. Finally, apply the same protocol to a product from another centre. The aim is transfer of the method, not memorisation of one provider’s parameter names.

Workshop 20: CF time bounds after aggregation

Begin with a hypothetical source package containing two versions of the same figure: an internal scientific chart and a public explanatory chart. The question is not whether both can share attractive wording, but whether their audiences need different layers of explanation without a different scientific claim. For the cf time bounds after aggregation case, first inspect the original record and identify which facts would survive a translation into a third language. Establish whether the data are an observation, model estimate, prediction, derived product or statistic; then list all units, times, spaces and version identifiers needed to reconstruct the figure.

Construct a comparison table with four columns: original machine field, source-language explanation, target-language explanation and independent QA evidence. Give special attention to total cloud cover because treating fractional cover as cloud thickness or rain probability changes variables. Cross-check subsurface runoff as a second variable; Preserve process and timescale. The exercise succeeds only when the technical distinctions are preserved while the target-language wording becomes easier to understand. An interface label can be brief if an accessible tooltip supplies the excluded details.

For the adversarial test, introduce one deliberate change in time support, coordinate reference, unit or ensemble meaning while leaving the visible numeric token unchanged. Ask another reader to determine whether the old caption is still true. A complete review records the mistake that was detected, the source evidence that resolved it, the approved wording, and the limit of what the chart can establish. Finally, apply the same protocol to a product from another centre. The aim is transfer of the method, not memorisation of one provider’s parameter names.

Workshop 21: Model calendar versus civil calendar

Begin with a hypothetical source package containing two versions of the same figure: an internal scientific chart and a public explanatory chart. The question is not whether both can share attractive wording, but whether their audiences need different layers of explanation without a different scientific claim. For the model calendar versus civil calendar case, first inspect the original record and identify which facts would survive a translation into a third language. Establish whether the data are an observation, model estimate, prediction, derived product or statistic; then list all units, times, spaces and version identifiers needed to reconstruct the figure.

Construct a comparison table with four columns: original machine field, source-language explanation, target-language explanation and independent QA evidence. Give special attention to total precipitation because replacing it with rain amount discards snow or other included frozen precipitation. Cross-check longwave radiation as a second variable; Keep radiation band and direction. The exercise succeeds only when the technical distinctions are preserved while the target-language wording becomes easier to understand. An interface label can be brief if an accessible tooltip supplies the excluded details.

For the adversarial test, introduce one deliberate change in time support, coordinate reference, unit or ensemble meaning while leaving the visible numeric token unchanged. Ask another reader to determine whether the old caption is still true. A complete review records the mistake that was detected, the source evidence that resolved it, the approved wording, and the limit of what the chart can establish. Finally, apply the same protocol to a product from another centre. The aim is transfer of the method, not memorisation of one provider’s parameter names.

Workshop 22: Negative flux and sign convention

Begin with a hypothetical source package containing two versions of the same figure: an internal scientific chart and a public explanatory chart. The question is not whether both can share attractive wording, but whether their audiences need different layers of explanation without a different scientific claim. For the negative flux and sign convention case, first inspect the original record and identify which facts would survive a translation into a third language. Establish whether the data are an observation, model estimate, prediction, derived product or statistic; then list all units, times, spaces and version identifiers needed to reconstruct the figure.

Construct a comparison table with four columns: original machine field, source-language explanation, target-language explanation and independent QA evidence. Give special attention to surface runoff because calling it observed river discharge changes location and unit. Cross-check cape as a second variable; Preserve diagnostic nature and units. The exercise succeeds only when the technical distinctions are preserved while the target-language wording becomes easier to understand. An interface label can be brief if an accessible tooltip supplies the excluded details.

For the adversarial test, introduce one deliberate change in time support, coordinate reference, unit or ensemble meaning while leaving the visible numeric token unchanged. Ask another reader to determine whether the old caption is still true. A complete review records the mistake that was detected, the source evidence that resolved it, the approved wording, and the limit of what the chart can establish. Finally, apply the same protocol to a product from another centre. The aim is transfer of the method, not memorisation of one provider’s parameter names.

Workshop 23: Bias correction and downscaling chain

Begin with a hypothetical source package containing two versions of the same figure: an internal scientific chart and a public explanatory chart. The question is not whether both can share attractive wording, but whether their audiences need different layers of explanation without a different scientific claim. For the bias correction and downscaling chain case, first inspect the original record and identify which facts would survive a translation into a third language. Establish whether the data are an observation, model estimate, prediction, derived product or statistic; then list all units, times, spaces and version identifiers needed to reconstruct the figure.

Construct a comparison table with four columns: original machine field, source-language explanation, target-language explanation and independent QA evidence. Give special attention to shortwave downward radiation because confusing J m−2 totals with W m−2 rates changes magnitude and duration. Cross-check ozone mixing ratio as a second variable; Preserve vertical level and composition basis. The exercise succeeds only when the technical distinctions are preserved while the target-language wording becomes easier to understand. An interface label can be brief if an accessible tooltip supplies the excluded details.

For the adversarial test, introduce one deliberate change in time support, coordinate reference, unit or ensemble meaning while leaving the visible numeric token unchanged. Ask another reader to determine whether the old caption is still true. A complete review records the mistake that was detected, the source evidence that resolved it, the approved wording, and the limit of what the chart can establish. Finally, apply the same protocol to a product from another centre. The aim is transfer of the method, not memorisation of one provider’s parameter names.

Workshop 24: Dataset DOI citation after a version update

Begin with a hypothetical source package containing two versions of the same figure: an internal scientific chart and a public explanatory chart. The question is not whether both can share attractive wording, but whether their audiences need different layers of explanation without a different scientific claim. For the dataset doi citation after a version update case, first inspect the original record and identify which facts would survive a translation into a third language. Establish whether the data are an observation, model estimate, prediction, derived product or statistic; then list all units, times, spaces and version identifiers needed to reconstruct the figure.

Construct a comparison table with four columns: original machine field, source-language explanation, target-language explanation and independent QA evidence. Give special attention to latent heat flux because treating it as a measured evaporation depth without latent heat conversion is invalid. Cross-check ensemble mean as a second variable; State aggregation set and statistic. The exercise succeeds only when the technical distinctions are preserved while the target-language wording becomes easier to understand. An interface label can be brief if an accessible tooltip supplies the excluded details.

For the adversarial test, introduce one deliberate change in time support, coordinate reference, unit or ensemble meaning while leaving the visible numeric token unchanged. Ask another reader to determine whether the old caption is still true. A complete review records the mistake that was detected, the source evidence that resolved it, the approved wording, and the limit of what the chart can establish. Finally, apply the same protocol to a product from another centre. The aim is transfer of the method, not memorisation of one provider’s parameter names.

A controlled terminology atlas: definition, safe label and a common trap

• 01. Analysis field. An estimated atmospheric state at an analysis time, often informed by data assimilation. Translator’s trap: Replacing estimated with directly observed erases modelling and analysis context. Editorial control: Write analysis estimate, state time, and keep its provenance separate from station observations.
• 02. Operational forecast. A model prediction tied to an initialization and a lead time. Translator’s trap: Calling a forecast an observation falsely upgrades evidence. Editorial control: Pair initialization with valid time and retain forecast designation.
• 03. Reanalysis product. A retrospective, model-assisted reconstruction using observations and a consistent production system. Translator’s trap: Calling reanalysis raw historical weather observations hides assimilation and uncertainty. Editorial control: Retain product name, release and reanalysis classification.
• 04. Hindcast. A retrospective model prediction or simulation used to evaluate behaviour over historical periods. Translator’s trap: Confusing hindcast and reanalysis loses how observations were used. Editorial control: Document simulation design and whether a run was initialized with historical inputs.
• 05. Nowcast. A very short-range estimate or forecast with product-specific methodology. Translator’s trap: Treating nowcast as a verified current observation overclaims certainty. Editorial control: Preserve product label and latency or forecast horizon.
• 06. Climate projection. A conditional future model result based on scenarios and assumptions. Translator’s trap: Calling a scenario-dependent projection a weather forecast suggests different predictive certainty. Editorial control: Preserve scenario, model, period and conditional language.
• 07. Climatological normal. A statistic calculated for a specified multi-year reference interval. Translator’s trap: Using normal to imply natural or desirable changes the claim. Editorial control: State statistic and baseline dates.
• 08. Climate anomaly. A departure relative to a stated reference climatology or baseline. Translator’s trap: Printing the departure without its baseline makes comparison ambiguous. Editorial control: Translate anomaly as difference from the named baseline.
• 09. Dataset edition. The identity of a delivered product and its particular published version. Translator’s trap: Mixing values or descriptions from two releases can manufacture apparent trends. Editorial control: Preserve version, release date and retrieval date.
• 10. Provisional release. A near-real-time release with a documented path to later replacement or adjustment. Translator’s trap: Suppressing provisional status portrays revision-prone data as final. Editorial control: Keep provisional qualifier and the expected revision pathway without inventing dates.
• 11. Analysis reference time. The timestamp associated with an analysis or initialization. Translator’s trap: Treating reference time as the time represented by any forecast step misplaces events. Editorial control: Store both clocks and their relation.
• 12. Forecast validity time. The time at which an instantaneous forecast applies or an interval field is indexed. Translator’s trap: Describing valid time as the start of any accumulation creates a shifted timeline. Editorial control: Display validity and interval bounds separately.
• 13. Forecast lead. Elapsed forecast time from the reference cycle, with units defined by the product. Translator’s trap: Assuming a bare 12 always means twelve hours ignores code-table time-unit choices. Editorial control: Decode unit and numerical lead before prose translation.
• 14. Production time. When a file was generated, assembled, or delivered. Translator’s trap: Confusing file creation with phenomenon time makes stale products appear current. Editorial control: Keep generation timestamp as distribution metadata.
• 15. Time-zone indication. A declared convention for interpreting timestamps. Translator’s trap: Adding an unqualified local-time label shifts the temporal referent. Editorial control: Preserve ISO timestamp and display locale conversion as an additional view.
• 16. Calendar attribute. The calendar model used by a netCDF time coordinate. Translator’s trap: Applying Gregorian arithmetic to a 360_day model calendar produces false dates. Editorial control: Keep calendar attribute and describe model-time conventions.
• 17. Time coordinate units. The numeric epoch and duration units used for time values. Translator’s trap: Reading hours since a base date as an absolute timestamp produces nonsense. Editorial control: Decode units and epoch together.
• 18. Instantaneous sample. A value representative of a particular time or instant. Translator’s trap: Calling a sample an hourly mean claims an averaging operation that never occurred. Editorial control: Use at time unless interval processing is documented.
• 19. Temporal mean. A statistic over a declared interval with beginning and end. Translator’s trap: Presenting mean temperature as a point observation conceals its time support. Editorial control: Preserve averaging method and interval bounds.
• 20. Temporal sum. A quantity accumulated or summed across a time interval. Translator’s trap: Confusing total depth with rate can inflate totals by hours or seconds. Editorial control: Report total, duration and interval endpoints.
• 21. Temporal maximum. The greatest reported value within a stated window. Translator’s trap: Writing maximum as instantaneous value misrepresents duration. Editorial control: Preserve window and method.
• 22. Temporal minimum. The lowest reported value within a stated window. Translator’s trap: Inferring the exact occurrence instant from window validity adds unsupported precision. Editorial control: Keep window and avoid inventing event time.
• 23. Rolling window. A statistic calculated for a window that moves between successive outputs. Translator’s trap: Summing overlapping rolling windows double-counts the same underlying hours. Editorial control: Mark window length, movement and overlap.
• 24. Accumulation reset. The point at which a run or product resets a cumulative quantity. Translator’s trap: Differencing across a reset may fabricate a negative rainfall amount. Editorial control: Check cycle boundaries before deriving increments.
• 25. Accumulation beginning. The first time in the temporal support of an accumulated field. Translator’s trap: Assuming beginning equals initialization loses partial-window products. Editorial control: Record beginning lead and ending lead.
• 26. Accumulation end. The endpoint that indexes a completed accumulation. Translator’s trap: Treating the endpoint as a single-instant observation changes the value class. Editorial control: Explain end time and duration.
• 27. Forecast step range. Two lead-time endpoints used to identify the supported interval. Translator’s trap: Translating only the latter number hides whether the statistic was accumulated from zero. Editorial control: Retain both step boundaries.
• 28. Accumulation versus rate. The distinction between total over a window and change per unit time. Translator’s trap: Adding rates as if they were totals ignores integration. Editorial control: State units, integration window and conversion assumptions.
• 29. Kelvin temperature. An absolute thermodynamic temperature with symbol K. Translator’s trap: Appending degrees or treating a change like an absolute temperature causes errors. Editorial control: Preserve absolute versus difference semantics.
• 30. Celsius display. A derived presentation of temperature from a compatible absolute source. Translator’s trap: Changing values without labelling the converted units breaks reproducibility. Editorial control: Keep native array and make display transform explicit.
• 31. Surface pressure. Atmospheric pressure at the model surface or terrain height. Translator’s trap: Calling surface pressure sea-level pressure disregards elevation and reduction method. Editorial control: Retain surface reference and Pa or hPa unit.
• 32. Mean sea-level pressure. Pressure adjusted to a reference at mean sea level using a defined method. Translator’s trap: Treating it as measured at ground level creates a false instrument comparison. Editorial control: Retain reduction concept.
• 33. Geopotential. Potential energy per unit mass associated with gravity in meteorology. Translator’s trap: Equating geopotential with geometric altitude without transformation can misstate vertical positions. Editorial control: Document source parameter and any converted product.
• 34. Pressure level. An isobaric vertical surface identified by pressure rather than geometric height. Translator’s trap: Interpreting 500 hPa as 500 metres is a categorical unit error. Editorial control: Preserve hPa label and variable’s vertical coordinate.
• 35. Hybrid model level. A terrain-following vertical coordinate defined by formula terms. Translator’s trap: Translating a model level number into metres invents altitude. Editorial control: Retain formula_terms and model-level identity.
• 36. Height above ground. An explicit vertical displacement relative to ground or a specified surface. Translator’s trap: Omitting the reference turns 2 m air temperature into an undefined vertical observation. Editorial control: Preserve height and datum.
• 37. Grid-cell mean. A quantity representative of a spatial cell or support, not a specific point sensor. Translator’s trap: Writing at a named city can imply exact local measurement. Editorial control: Use over the grid cell and note spatial resolution.
• 38. Grid-point value. A quantity evaluated at a discrete grid coordinate with model support. Translator’s trap: Describing one point as a regional average can change representativeness. Editorial control: Record interpolation and grid definition.
• 39. Latitude ordering. Whether array rows run north to south or south to north. Translator’s trap: Reversing row order while keeping latitude labels unchanged flips a map vertically. Editorial control: Validate coordinate monotonicity and bounds.
• 40. Longitude convention. Whether longitudes are represented as zero to 360 or minus 180 to 180. Translator’s trap: Treating an equivalent wrapped longitude as a different physical position duplicates points. Editorial control: Normalize only for display with reversible mapping.
• 41. Rotated grid. A coordinate grid whose axes differ from ordinary geodetic latitude and longitude. Translator’s trap: Labelling rotated coordinates directly as Earth latitude and longitude shifts locations. Editorial control: Preserve rotated coordinate definition.
• 42. Projected grid. A coordinate system defined by map projection and its parameters. Translator’s trap: Relabelling projected x/y as latitude/longitude destroys geographic meaning. Editorial control: Carry projection metadata and coordinates.
• 43. Grid mapping. CF metadata connecting native coordinates to a coordinate reference system. Translator’s trap: Deleting the mapping can make values appear in a different region. Editorial control: Retain grid_mapping and a separately localized description.
• 44. Cell bounds. Coordinates delineating a cell’s interval or spatial area. Translator’s trap: Assuming a point coordinate always means the centre of a rectangular cell introduces unsupported geometry. Editorial control: Preserve bounds and do not invent centroids.
• 45. Cell area. Area used for spatial weighting or integration. Translator’s trap: Averaging unequal-area cells without weights can bias a global statistic. Editorial control: Retain cell_measures or appropriate area weights.
• 46. Land-sea mask. A field identifying land/ocean fraction or classification in a cell. Translator’s trap: Treating partial land fraction as an absolute coastline invents resolution. Editorial control: Explain fractional or categorical mask and grid support.
• 47. Oroography. Representation of surface elevation in a model grid. Translator’s trap: Describing model surface altitude as precise local topography falsely sharpens the data. Editorial control: Label terrain representation and source resolution.
• 48. Interpolation method. A transformation estimating values onto new coordinates. Translator’s trap: Calling regridded numbers native measurements conceals a derived calculation. Editorial control: Record source grid, target grid and method.
• 49. Regridding conservativeness. Whether a spatial transform conserves an extensive quantity under stated assumptions. Translator’s trap: Applying a generic interpolator to area totals may change integrated quantities. Editorial control: Keep method and check integrated totals.
• 50. Wind U component. The eastward or grid-relative component of wind velocity. Translator’s trap: Rendering U as wind speed ignores vector direction and possible grid rotation. Editorial control: State component orientation and sign.
• 51. Wind V component. The northward or grid-relative wind component. Translator’s trap: Replacing V with a plain speed loses vector orientation. Editorial control: Keep coordinate basis and component label.
• 52. Wind speed magnitude. The magnitude calculated from vector components or supplied by a model. Translator’s trap: Taking arithmetic means of magnitudes as vector-mean wind changes statistics. Editorial control: Preserve magnitude and aggregation method.
• 53. Wind direction convention. The convention for direction from which or toward which wind travels. Translator’s trap: Flipping from-direction to to-direction changes meaning by 180 degrees. Editorial control: Declare convention and angle reference.
• 54. Wind gust. A short-duration peak under a model or observational definition. Translator’s trap: Treating a gust as mean wind speed exaggerates sustained conditions. Editorial control: Preserve peak definition and window.
• 55. Dew point. Temperature at which air reaches saturation for a given moisture state. Translator’s trap: Mistaking dew point for air temperature changes humidity interpretation. Editorial control: Retain physical definition.
• 56. Specific humidity. Mass of water vapour per mass of moist air. Translator’s trap: Confusing specific humidity and mixing ratio changes denominator. Editorial control: Preserve exact property and unit.
• 57. Relative humidity. Vapour pressure relative to saturation under a defined temperature basis. Translator’s trap: Treating percent relative humidity as mass percentage creates a false material balance. Editorial control: Keep saturation-based definition and percent unit.
• 58. Total cloud cover. Fraction of sky or grid area represented by cloud cover under a product definition. Translator’s trap: Treating fractional cover as cloud thickness or rain probability changes variables. Editorial control: Label coverage and support.
• 59. Convective precipitation. Precipitation parameter attributed to a model’s convective scheme. Translator’s trap: Translating the category as all thunderstorm rainfall may overstate physical attribution. Editorial control: Retain parameter distinction.
• 60. Large-scale precipitation. Precipitation attributed to resolved cloud or large-scale processes in the model. Translator’s trap: Equating large-scale with wide geographic extent is too literal. Editorial control: Explain model process classification.
• 61. Total precipitation. Water-equivalent precipitation amount over a defined interval. Translator’s trap: Replacing it with rain amount discards snow or other included frozen precipitation. Editorial control: Keep inclusions, units and accumulation bounds.
• 62. Snowfall water equivalent. Liquid-water equivalent corresponding to a snow quantity. Translator’s trap: Equating it directly with snow depth ignores density. Editorial control: Identify water-equivalent basis.
• 63. Snow depth. The modelled or measured vertical depth of snow cover. Translator’s trap: Misreading depth as water equivalent breaks storage estimates. Editorial control: Keep physical depth and unit.
• 64. Surface runoff. Water that leaves the land surface as flow under a model’s water balance. Translator’s trap: Calling it observed river discharge changes location and unit. Editorial control: Report depth and spatial basis.
• 65. Subsurface runoff. Modelled below-surface drainage component over a given time. Translator’s trap: Calling all subsurface runoff groundwater recharge ignores routing and definitions. Editorial control: Preserve process and timescale.
• 66. Soil moisture layer. Water content stored in a specified soil depth or model layer. Translator’s trap: Conflating volumetric water content with depth-integrated water storage changes units. Editorial control: Preserve layer bounds and water-content basis.
• 67. Shortwave downward radiation. Solar-origin radiation incident on a surface, expressed as flux or accumulated energy. Translator’s trap: Confusing J m−2 totals with W m−2 rates changes magnitude and duration. Editorial control: Report direction, unit and interval.
• 68. Longwave radiation. Thermal infrared radiative flux or accumulated energy with directional specification. Translator’s trap: Removing upward or downward can reverse an energy-budget interpretation. Editorial control: Keep radiation band and direction.
• 69. Sensible heat flux. Turbulent transfer of heat without phase-change latent energy. Translator’s trap: Dropping sign convention can reverse surface heating interpretation. Editorial control: Check positive-direction rule.
• 70. Latent heat flux. Energy associated with moisture phase change and turbulent exchange. Translator’s trap: Treating it as a measured evaporation depth without latent heat conversion is invalid. Editorial control: Preserve energy form and sign.
• 71. CAPE. Convective available potential energy of an atmospheric parcel under specified assumptions. Translator’s trap: Calling CAPE a probability of storms is a category error. Editorial control: Preserve diagnostic nature and units.
• 72. Boundary-layer height. A model-derived estimate of atmospheric boundary-layer depth. Translator’s trap: Reporting it as the altitude of every cloud base invents a relationship. Editorial control: Retain diagnostic definition.
• 73. Visibility. A visibility diagnostic or observation with model-dependent definition. Translator’s trap: Interpreting visual range as proof of clear air under all conditions overclaims. Editorial control: Specify method and unit.
• 74. Ozone mixing ratio. An atmospheric composition quantity with a specific concentration or mixing-ratio unit. Translator’s trap: Treating all ozone fields as surface air quality changes altitude and relevance. Editorial control: Preserve vertical level and composition basis.
• 75. Ensemble control. A designated member produced with the system’s control configuration. Translator’s trap: Calling the control member the most likely outcome misstates ensemble interpretation. Editorial control: Keep member identity and generation description.
• 76. Perturbed member. An ensemble realization with an intentional perturbation in initial conditions or model representation. Translator’s trap: Treating member number as ranking invents an order. Editorial control: Preserve member number without probability claims.
• 77. Ensemble mean. A statistic across selected members at a specified time and variable. Translator’s trap: Treating the mean as an actual member’s trajectory may create impossible combinations. Editorial control: State aggregation set and statistic.
• 78. Ensemble spread. A dispersion statistic whose definition may be standard deviation or another measure. Translator’s trap: Translating spread as error or confidence interval without method overstates inference. Editorial control: Name the statistic and sample population.
• 79. Ensemble percentile. A quantile of a specified ensemble distribution. Translator’s trap: Confusing 90th percentile with 90 percent event probability changes assertion. Editorial control: Preserve percentile definition.
• 80. Threshold probability. Share or modeled probability of members exceeding a declared event threshold. Translator’s trap: Suppressing threshold makes a probability meaningless. Editorial control: State event threshold, interval and member basis.
• 81. Ensemble count. Number of included, valid ensemble members in a statistic. Translator’s trap: Hiding missing members can suggest comparability where sample sizes differ. Editorial control: Report count and exclusions.
• 82. Deterministic versus probabilistic products. Whether output is a single scenario or a distributional summary. Translator’s trap: Calling every forecast probabilistic implies calibrated likelihoods that may not exist. Editorial control: Keep product classification.
• 83. Bias correction. A statistical transformation intended to adjust systematic model differences. Translator’s trap: Calling corrected data direct observations hides a derived product. Editorial control: Document training period, method and residual limits.
• 84. Downscaling. A method for deriving finer-resolution fields using dynamics or statistics. Translator’s trap: Calling downscaled cells new sensor measurements creates false precision. Editorial control: State method and source resolution.
• 85. Extreme-event attribution. A structured estimation of how conditions influence event likelihood or intensity. Translator’s trap: Treating attribution as proof that one event had a single cause misstates uncertainty. Editorial control: Preserve conditional phrasing and uncertainty.
• 86. GRIB2 discipline and parameter. A coded identification system for the type of physical parameter. Translator’s trap: Translating a parameter label without discipline and category risks a code collision. Editorial control: Keep code-table path with descriptive label.
• 87. GRIB2 product definition template. A structured template describing forecast type and statistical processing. Translator’s trap: Ignoring product definition loses ensemble or time-window semantics. Editorial control: Record template and decoded fields.
• 88. GRIB2 grid definition. A coded description of the spatial grid and scanning order. Translator’s trap: Translating grid labels while dropping scanning direction can misplace arrays. Editorial control: Preserve grid definition and scan mode.
• 89. GRIB2 local tables. Codes assigned by a originating centre in a locally scoped table. Translator’s trap: Treating local code values as universal WMO definitions introduces wrong parameters. Editorial control: Record producer and local-table version.
• 90. GRIB2 packing and precision. Methods that encode numerical arrays with specified scaling and representation. Translator’s trap: Assuming stored integers are the physical values changes units and range. Editorial control: Decode before translation or arithmetic.
• 91. GRIB2 bitmap. A mask indicating availability or missing grid-cell values. Translator’s trap: Rendering absent cells as zeros invents meteorology. Editorial control: Preserve bitmap and null semantics.
• 92. CF standard_name. A controlled identifier for the physical quantity of a netCDF variable. Translator’s trap: Translating standard_name itself can break CF interoperability. Editorial control: Separate machine identifiers and prose descriptions.
• 93. CF long_name. A free-text human-readable description of a variable. Translator’s trap: Treating long_name as an authority over conflicting units and coordinates conceals errors. Editorial control: Cross-check all metadata fields.
• 94. CF units. A unit-string attribute interpreted with recognized unit conventions. Translator’s trap: Replacing native unit strings with nonstandard translated words may break software. Editorial control: Keep units parseable and translations outside the attribute.
• 95. CF cell_methods. A structured attribute identifying aggregation over particular axes. Translator’s trap: Localizing keywords inside the attribute destroys machine semantics. Editorial control: Preserve controlled tokens and localized explanation.
• 96. CF time bounds. A boundary variable giving the start and end of time intervals. Translator’s trap: Discarding bounds makes means and sums harder to interpret. Editorial control: Retain boundary variable and verify convention.
• 97. CF flag_values and flag_meanings. Machine-readable associations between values and defined categories. Translator’s trap: Translating flag_meanings tokens in place may damage downstream parsing. Editorial control: Preserve mappings and provide an external legend.
• 98. CF history metadata. A free-text record of processing steps where available. Translator’s trap: Deleting preprocessing history hides regridding, merges and conversions. Editorial control: Retain processing trace and append a transparent localization note.
• 99. Dataset citation and DOI. A stable bibliographic reference for a specific data product or publication. Translator’s trap: Changing citation while retaining data makes provenance unrecoverable. Editorial control: Preserve exact DOI and access context.
• 100. Checksum and integrity. A digest or other mechanism for confirming file-byte identity. Translator’s trap: Describing a transformed file as byte-identical misrepresents integrity checks. Editorial control: Keep separate integrity records for source and derivative.

Thirty questions frequently asked about weather and reanalysis translation

FAQ 01: Can GRIB2 be translated like ordinary prose?

No. Translate surrounding labels and documentation; preserve or deliberately map controlled numeric codes and structures using the applicable WMO and centre tables. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 02: Can I change a CF standard_name into another language?

Keep the canonical machine token unchanged. Put a localized explanation in an appropriate separate human-language field or interface string. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 03: Does ERA5 mean a historical observation?

ERA5 is a retrospective reanalysis, integrating observations through a model and assimilation system. It should not be labelled a direct single-station observation. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 04: Are forecast reference time and valid time the same?

Not generally. For a forecast they differ by the decoded forecast step; interval fields also need start and end support. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 05: Can I use local time to help a reader?

Yes as an additional calculated display, but preserve UTC reference and make daylight-saving or calendar conventions explicit when relevant. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 06: Does a value stamped 18 UTC describe the previous hour?

It depends on the variable and product. Examine temporal processing, bounds and source documentation rather than assuming. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 07: Is total precipitation measured in millimetres?

Some products store metres of water equivalent and others use other units. Read the native unit; convert explicitly if necessary. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 08: Does a one-hour precipitation total equal a precipitation rate?

No. They are different quantities. A rate can be integrated over a known duration only with correct temporal and unit assumptions. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 09: Does snowfall water equivalent describe snow depth?

No. Physical snow depth also depends on density and snowpack structure. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 10: Is Kelvin the same as Celsius?

Absolute readings differ by 273.15 in conventional conversion; temperature differences have the same increment size. Keep that distinction. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 11: Can Pa be replaced by hPa without converting the number?

No. One hectopascal equals 100 pascals, so the numeric label must change when units change. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 12: Does 500 hPa mean 500 metres above ground?

No. It specifies an atmospheric pressure surface, whose geometric height is variable. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 13: Does a model layer number identify an altitude?

No. It is a coordinate index or value whose spatial interpretation may require formula terms and other fields. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 14: Does higher horizontal resolution guarantee precise local weather?

No. Grid spacing is not the same as observational support or verified accuracy at a particular site. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 15: Can longitude 350 degrees east be normalized?

Yes. It is equivalent to minus ten degrees in a signed-longitude convention, provided the mapping is handled consistently. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 16: Can netCDF conversion silently change meanings?

Yes if variables, dimensions, calendars, bounds, processing and coordinates are dropped or mis-mapped. Validate before and after. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 17: What does cell_methods protect?

It distinguishes how data were summarized over dimensions, such as means, sums and maxima. Controlled method tokens are not free translation text. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 18: Why do cell bounds matter?

They identify the intervals or spatial cells represented by values, often essential to interpreting means and accumulations. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 19: What is the difference between U wind and wind speed?

U is a signed component along a specified axis; wind speed is a magnitude, normally requiring at least two components in a horizontal vector. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 20: Does a wind direction state where the wind is going?

Meteorological wind direction commonly names the direction from which it blows. Check the actual source convention. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 21: Is ensemble spread an error bar?

Not automatically. It is a measure of dispersion across ensemble products; its definition and calibration matter. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 22: Is a 90th percentile the same as a 90 percent chance?

No. A quantile value and a probability of exceeding a threshold answer different questions. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 23: Why retain the ensemble member count?

Statistics can change meaning and stability when members are excluded or the number of members differs. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 24: Can an anomaly be translated without naming a baseline?

It becomes less interpretable and may be misleading. State the reference period and statistic. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 25: What if the WMO code list is updated?

Record the version applicable to the file. A code’s meaning should not be retroactively reassigned from an unrelated future table. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 26: Can I translate a local GRIB code by looking up its number only?

No. Resolve the generating centre and local-table scope as well as code identifiers. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 27: Is a masked cell the same as zero?

No. Missing, outside-domain, and zero are distinct data states. Preserve the mask and field semantics. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 28: What qualifies as good bilingual QA?

Compare the translated claim against the data and metadata, evaluate both scientific accuracy and readability, and document defects by severity. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 29: Should a data-translating model be trusted without review?

No. Use it as an editorial aid at most; check times, units, identifiers, negation, qualifiers and method distinctions against authoritative sources. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

FAQ 30: Where should readers go after this specialist system?

Return to the Master Art of Translation root for general meaning-transfer practice and to the protected language-learning hubs only for their respective language skills. The practical way to verify this answer is to inspect an actual source file, identify which field fixes the interpretation, then test a nearby but different product. Whenever the value depends on a source-specific assumption, the translated statement should name the assumption and preserve an audit path. The goal is a sentence that someone can use correctly, not merely recognize as familiar terminology.

Four-week implementation plan for a multilingual scientific data team

Week one: evidence inventory and basic temporal literacy

Select one GRIB2 file and one CF-netCDF output of a documented product. Catalogue all variables, clocks, bounds, units and references. Write a source-to-target glossary but translate no machine identifiers. A second person must be able to reproduce the basic metadata inventory. Document acceptance criteria for each activity: source fields identified, invariant meanings preserved, numerical derivations tested, unresolved dependencies recorded and independent reconstruction passed. A completed lesson should leave a reusable QA artefact, not merely a folder of translated paragraphs.

Week two: conversions and falsification tests

Practice reversible time, coordinate and unit displays on toy records. Introduce deliberately wrong accumulation windows, missing-value replacements and vertical-reference labels. Record how each would be detected and which tests distinguish it from a valid alternative. Document acceptance criteria for each activity: source fields identified, invariant meanings preserved, numerical derivations tested, unresolved dependencies recorded and independent reconstruction passed. A completed lesson should leave a reusable QA artefact, not merely a folder of translated paragraphs.

Week three: ensemble and public communication

Create captions for a deterministic forecast, an ensemble mean, a percentile and a threshold probability. Let non-specialists read the captions and write down the quantity each believes is being described. Repair every ambiguity that yields a false interpretation. Document acceptance criteria for each activity: source fields identified, invariant meanings preserved, numerical derivations tested, unresolved dependencies recorded and independent reconstruction passed. A completed lesson should leave a reusable QA artefact, not merely a folder of translated paragraphs.

Week four: governance and publication

Publish a controlled demonstration with dataset citation, licence, source snapshot, version, QA record and a clear fictional-data label. Audit accessibility, cross-language labels and mobile charts. Define who rechecks content when the source dataset or code table changes. Document acceptance criteria for each activity: source fields identified, invariant meanings preserved, numerical derivations tested, unresolved dependencies recorded and independent reconstruction passed. A completed lesson should leave a reusable QA artefact, not merely a folder of translated paragraphs.

Quality rubric: how to evaluate whether meaning actually survived

Evaluate source fidelity separately from style. A critical error changes a physical variable, shifts the clock or interval, moves a grid cell, changes a sign or order-of-magnitude unit, destroys a controlled identifier, or turns forecast uncertainty into certainty. A major error obscures an important qualifier, processing method or source release. A moderate error weakens transparency but leaves a careful expert able to reconstruct the claim. Minor issues concern terminology consistency or readability without scientific impact. This classification is an editorial policy for this article, not a claim to reproduce the proprietary text of any standard.

Sampling should reflect risk. Check all high-consequence numbers and commands, not merely a random subset of captions. A long dataset dictionary may need representative sampling of routine low-risk prose while preserving complete structural comparison for identifiers, units, times, bounds and coordinate definitions. Document the sample logic, number of items reviewed and defect category. After correcting one error, search for the same conceptual error across parallel languages and related variables.

No single metric can prove a translation correct. Reference-based text similarity can reward fluent but wrong field descriptions, and a structural validator can accept numerically correct values embedded in misleading prose. Combine bilingual domain expertise, deterministic schema checks, reconstructed reader meaning and source provenance. Treat machine translation as an aid to drafts, and subject any generated explanation of scientific data to explicit verification.

Sources, standards and careful limits

• WMO Manual on Codes, Volume I.2: machine-readable GRIB2 edition 37, valid 1 June 2026. For coded GRIB2 parameter and template interpretation.
• WMO Codes Registry. For registered meteorological code lists and scope.
• CF Metadata Conventions 1.13, published 17 December 2025. For variable definitions, units, coordinates, bounds, and cell methods.
• ECMWF ERA5 product overview. For the identity and scope of ERA5 reanalysis.
• Copernicus Climate Change Service: climate reanalysis. For reanalysis terminology, uncertainty and provisional data.
• ECMWF ERA5 post-processed daily statistics catalogue. For transparent daily aggregation options and provenance.
• NOAA Climate Prediction Center: wgrib2 forecast time units. For examples of forecast lead-time encoding.
• ISO 5060:2024 translation output evaluation guidance. For translation-evaluation principles and sampling scope.

The official source-specific product documentation is always decisive for any downloaded file. The training scenarios here are intentionally invented so readers can practise mechanisms without mistaking them for an actual forecast, weather warning or dataset observation. No invented code number, coordinate or measurement should be copied into a production system.

Next routes: return to translation, then choose the correct protected specialist

The main return route is Master Art of Translation | The Complete System for Moving Meaning Between Languages. This master system explains how to preserve claims across languages, genres and evidence types; the present child applies those principles to atmospheric data formats and scientific metadata. For a deeper study of how readers acquire and distinguish technical words, use the protected Vocabulary Learning Hub without treating this article as a competing vocabulary owner. For the grammar of qualified claims, sentence-level ambiguity, hedging, source attribution or scientific explanation in English, use the protected How English Works | The Complete Map and Learning Routes.

The immediate task is to take one legitimate source file, save its exact snapshot, write a semantic envelope, preserve code-facing identifiers, translate only the reader-facing words, pass the release gates and test whether an independent person reconstructs the same physical claim. Successful translation retains the relationship between measurement, model, space, time, uncertainty and source. Those relationships are the meaning that must cross the language boundary.