Master Art of Translation | The OpenLineage Data Lineage and Pipeline Metadata Translation System — How Runs, Jobs, Dataset Facets, Column Dependencies and Provenance Keep Their Meaning

How do you translate OpenLineage data lineage metadata, pipeline documentation, dataset descriptions and column-dependency explanations into another language without breaking jobs, runs, facets or data provenance? The safe answer is to localize human-facing descriptions and interface labels while preserving the event schema, identifiers, namespaces, run states, timestamps, dataset and field identities, lineage edges and quality evidence. This complete guide shows how to translate a data lineage catalog accurately, review multilingual pipeline documentation, and validate the result against real machine-readable OpenLineage events.

The hidden problem in multilingual data governance is that a natural-sounding translation may silently change what a lineage graph claims. A localized job name can become a different technical job, an input column can be described as an output, a scheduled time can be confused with the actual event time, and an indirect filtering dependency can be presented as a direct transfer of values. A translation that makes a dashboard easier to read but changes who produced what, when, or from which dataset is not a successful translation. It is a broken evidence trail.

Quick answer and method: first identify the exact RunEvent, JobEvent or DatasetEvent; lock all machine identifiers, enumerations, schema versions, data types, event times and provenance links; classify each description as translatable prose, immutable evidence or controlled terminology; translate explanatory labels in a separate, versioned display layer; review business meaning with a domain owner; compare the original and localized dependency graphs; test representative runs and failures; and release only when the same technical facts appear in every language. This world-facing textbook develops that repeatable system through field-by-field diagnostics, failure analysis and transfer workshops.

• Reader: translators, data engineers, data stewards, catalog owners, localization teams, analysts and technical writers working on multilingual metadata.
• Search problem solved: how to localize OpenLineage events, dataset names, column-lineage descriptions, lineage graphs, data quality reports and pipeline incident notices safely.
• Not the task: rewriting live source tables, changing event payload IDs, renaming columns, editing an orchestrator, or modifying the lineage protocol.
• Deliverable: a bilingual or multilingual display dictionary, protected-field inventory, semantic review log, graph-equivalence report and approved release bundle.
• Evidence standard: a translation is accepted only when both human interpretation and machine-verifiable relationships remain correct.

Search intent: what does translating OpenLineage metadata actually mean?

A reader searching for “translate OpenLineage documentation”, “localize data lineage metadata”, “multilingual data catalog”, “column lineage translation”, “OpenLineage facets”, or “how to translate pipeline descriptions” is usually solving two problems at once. The reader wants familiar language on a screen, but the screen is reporting facts from technical records. Those technical records have identities, versioned schemas, event states and dependencies that must not be paraphrased into different data. A useful solution therefore separates a language layer from the evidence layer and records their relationship explicitly.

OpenLineage is an open-source specification for collecting lineage information about data jobs and datasets. Its object model describes Jobs, Runs and Datasets, and its events communicate metadata about their activity. Facets add schema, ownership, quality, documentation and other context. This is why a translator must work at a finer level than sentences: the prose inside a documentation facet is a different kind of object from the Job identity or a transformation type. The first may be localized; the second must ordinarily remain stable. See the official OpenLineage object model and facets documentation.

This article does not claim that OpenLineage has a built-in general multilingual display standard. The recommended display dictionary, approval process and localized glosses below are editorial and engineering design patterns, not additional required OpenLineage fields. Implementations may vary. When a field is described as protected, that is a translation-safety rule for the underlying contract, not a claim that every identifier is technically immutable forever: an authorized schema migration is a separate change requiring its own audit and rollout.

The hidden problem: translation can change a lineage graph without anyone moving a single row

Imagine a lineage viewer showing “Customer Status”, “Active Accounts”, “Source Orders” and “Daily Revenue”. An editor might translate every visible phrase into idiomatic language. But those strings may come from different layers: an actual column identifier, a field description, a business glossary definition, an application menu label and the human title of a dataset. If the editor conflates them, an apparently successful translation may make two separate columns look identical or make a single dataset appear to have two histories. Users may then make incorrect incident, audit or regulatory decisions even though the underlying event feed remains untouched.

The first invisible distinction is identity versus explanation. A namespace and name pair identifies a Job or Dataset inside the lineage model. A localized alias may help a reader but must continue to point at the same pair. The second distinction is observation versus interpretation. A metric can report a null count of 42; a translator may describe it as “42 missing values” if the metric semantics support that description, but must not call it 42 failed transactions. The third distinction is direct versus indirect influence. A field used only in a WHERE filter can influence which rows survive without supplying the output field’s value.

The fourth distinction is event time versus business time. A pipeline scheduled for midnight may actually execute later; the date carried by an invoice may be earlier still. The fifth is declared versus observed lineage. A curated relationship in a catalog is not automatically a dependency proved during a particular run. Finally, machine-readable evidence versus user-facing narration must stay visibly separate. Translating an incident notice requires a different kind of permission from editing event data.

The mechanism: five layers of meaning

Layer one: immutable reference points

Record event IDs, run IDs, Job namespace and name, Dataset namespace and name, field keys, schema pointers, producer URI, event types and reference timestamps. Their exact bytes often determine whether systems can join or interpret records. Translate their explanations, not the tokens. A useful reviewer asks not only whether the sentence is grammatical, but what evidence makes the sentence true and where a future investigator can verify that evidence.

Layer two: typed observations

Interpret whether a numeric value is a count, timestamp, byte quantity, duration, measurement, version, error code or identifier. Before translating its nearby words, write down its unit, scope and collection time; these properties prevent misleadingly fluent paraphrases. A useful reviewer asks not only whether the sentence is grammatical, but what evidence makes the sentence true and where a future investigator can verify that evidence.

Layer three: causal relationships

Distinguish input and output datasets, table lineage and field lineage, direct value derivation and indirect selection effects. A relationship can be true without appearing as a word in a field title. The graph is a separate artifact to validate. A useful reviewer asks not only whether the sentence is grammatical, but what evidence makes the sentence true and where a future investigator can verify that evidence.

Layer four: human language

Locate documentation descriptions, field descriptions, tags intended for presentation, catalog glossaries, dashboard headings, error explanations and release notices. Translate these using a stable terminology register with examples and warnings. Do not assume every JSON string is prose. A useful reviewer asks not only whether the sentence is grammatical, but what evidence makes the sentence true and where a future investigator can verify that evidence.

Layer five: authorization and change history

Identify who approved the source definition, who translated it, the target audience and the version it describes. Preserve source provenance and keep translations editable without rewriting historical events. If a meaning changes, record a new interpretation or authorized migration instead of silently overwriting history. A useful reviewer asks not only whether the sentence is grammatical, but what evidence makes the sentence true and where a future investigator can verify that evidence.

A small fictional event: technical data and a separate human display

{
  "eventType": "COMPLETE",
  "eventTime": "2026-10-01T04:05:00Z",
  "producer": "https://example.org/lineage-emitter/2.1",
  "schemaURL": "https://openlineage.io/spec/2-0-2/OpenLineage.json#/$defs/RunEvent",
  "run": {
    "runId": "d2b523f4-57cd-4d2a-80e8-5eabc8598ff5"
  },
  "job": {
    "namespace": "tutorial-workflows",
    "name": "refresh_daily_invoices"
  },
  "inputs": [
    {
      "namespace": "postgresql://example.org:5432",
      "name": "raw.invoices"
    }
  ],
  "outputs": [
    {
      "namespace": "postgresql://example.org:5432",
      "name": "analytics.daily_invoice_summary"
    }
  ]
}

This synthetic example is explanatory, not a claim about an actual pipeline. A target-language reader may see a heading such as “Daily invoice summary refreshed” and a translated description of the calculation, but the eventType remains COMPLETE; the runId remains identical; the datasets remain the same; and the eventTime remains an unambiguous timestamp. If the title is translated into several languages, each translated view must still resolve to this one event and the same lineage identities. The example points to the official run-cycle model, which distinguishes START, RUNNING, COMPLETE, ABORT, FAIL and OTHER.

Repeatable method: a 28-gate translation and verification process

01. Identify the reader’s decision

Write down whether the localized text will guide incident diagnosis, business reporting, data-discovery search or developer debugging. A phrase that is safe for a marketing card may be dangerously vague in an audit view. Define the decision before the language. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

02. Freeze a representative evidence set

Export read-only examples of a successful run, a failed run, a resumed run, an unchanged dataset and a changed dataset. Capture their authoritative event versions and retrieval timestamps. Do not test only the happy path. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

03. Inventory protocol fields

List the keys and values used for eventType, eventTime, producer, schemaURL, Run, Job, Dataset, fields and facets. Mark whether a field is an identifier, enum, quantity, free text, code or controlled vocabulary. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

04. Identify the source of each displayed string

Trace every visible label back to its origin: OpenLineage event, catalog database, orchestrator configuration, application UI, business glossary or hand-written runbook. A display alias without provenance cannot be maintained reliably. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

05. Set protected-field boundaries

Keep runId, namespace, machine name, schema pointer, event-type token, SQL query and graph-edge identifiers outside the translation memory’s replacement scope. Record approved exceptions only through engineering change control. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

06. Build a bilingual field dictionary

For each safe-to-translate phrase, store source text, target language, chosen wording, source field path, translator, reviewer, context and current status. Include disambiguation notes for ambiguous terms such as run, job, field, source and output. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

07. Define units and business meaning

Annotate counts, bytes, dates, currencies, success rates and missing-value statistics with their exact measurement definitions. Translation must not supply a unit, interpretation or data-quality judgment the source has not established. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

08. Protect the event lifecycle

Preserve the six eventType tokens. Write approved target-language labels for status displays rather than translating enumerations in transmitted JSON. Verify terminal versus nonterminal behavior with a run lifecycle example. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

09. Protect graph direction

Verify that each input points into the same transformation or job and each output points out. Sketch one graph before translation and compare its edge list after language changes. Human labels must not reverse causal direction. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

10. Inspect field-level dependencies

For each output field, document which source fields directly contribute values and which indirectly affect selection, grouping or ordering. Distinguish the semantics of DIRECT and INDIRECT before choosing verbs. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

11. Separate expected and observed time

Document eventTime, nominal time and any business date used in reporting. Show time zones when the UI requires local conversion; do not convert a scheduling target into a historical fact. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

12. Separate producer from owner

The producer URI identifies the source of lineage metadata; an ownership facet can identify an accountable party. Do not translate either into “the person who ran this job” without evidence. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

13. Confirm target-language technical terminology

Ask domain reviewers how they distinguish run, pipeline, task, input, upstream dependency, dataset, field, column, freshness and lineage. Maintain distinctions even when two words translate similarly. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

14. Handle constrained user interfaces

Test short labels, breadcrumbs, graph nodes, hover tooltips, tables, screen-reader text and exported PDFs. When two long translated labels truncate identically, add identifiers or disambiguators rather than hiding the distinction. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

15. Review mixed-language payloads

Inspect code fragments, quoted names, escaped characters, SQL literals, file paths, URIs and schema pointers. A sentence can legitimately contain preserved English identifiers alongside localized prose. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

16. Test missing descriptions

Plan a visible fallback for untranslated or unknown-language text. A fallback must not masquerade as a completed translation; preserve source text with an explicit language or coverage indicator. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

17. Audit incident terminology

Use genuine run evidence before calling a task complete, failed, aborted, retried or skipped. Different orchestration tools use different status words; the OpenLineage event enum is the reference point for the display mapping. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

18. Audit schema meaning

Compare field names, data types and nested structure. Localized descriptions are not permission to rename underlying fields, infer units or flatten nested records. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

19. Audit lineage extraction limitations

Find out whether edges are observed, statically inferred, manually curated or incomplete. Do not translate “possible relationship” into “guaranteed data transfer” in a catalog description. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

20. Audit data quality claims

Link every displayed statistic to captureDate or comparable evidence, source dataset and field. Distinguish null counts, row counts and assertions; avoid making stale measurements sound live. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

21. Validate interface behavior

Run deterministic test fixtures through the localized user interface. Confirm sorting, search, graph selection, deep links and filters resolve to identical technical entities in the source and target languages. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

22. Validate exported records

If the user can export localized reports, verify that IDs, code values, event times, data-lineage edges and source references are preserved alongside the translated prose. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

23. Review accessibility

Write meaningful alternative text for graphs, clear heading structure and expanded abbreviations where appropriate. Ensure screen readers can distinguish visual aliases from exact source identifiers. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

24. Run a red-team terminology pass

Ask a reviewer to find where a translation could cause an analyst to investigate the wrong dataset, a maintainer to roll back the wrong job or an auditor to mistake a preliminary metric for proof. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

25. Record each unresolved ambiguity

Preserve competing interpretations and the exact evidence needed to resolve them. Never silently choose a fluent target phrase where the source lineage is ambiguous. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

26. Approve through two independent lenses

A language reviewer checks fluency, register and consistency. A data-domain reviewer checks technical truth, graph relationships, quantities and evidence. Both must sign off on high-consequence claims. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

27. Release with reversible mapping

Publish translated display dictionaries by version and language. Keep a way to recover the original description and to roll back a defective localization without restoring or rewriting historical run data. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

28. Monitor reader errors

After launch, watch searches returning wrong entities, misclicks on similarly named jobs, incident messages misunderstood by users and translation change requests. Feed these findings into the next reviewed dictionary version. Exit evidence: the reviewer can show a concrete example, a source pointer and a yes/no verification result rather than merely reporting that the translation “looks good.” The next gate inherits this evidence so defects have a precise place to be caught and corrected.

Failure modes: what to reject before release

Failure 1: Rewriting a namespace

A natural-language translation of an orchestrator namespace creates a second identity instead of a second way of naming the first one. Reject the output and restore the exact namespace in every structured event. Practical diagnostic: identify the source object; inspect its field path and type; compare the exact source event with the localized display; ask whether any decision-critical value or dependency changed. If it did, block release and correct the display mapping, not the source-of-truth record.

Failure 2: Localizing enumerations inside event data

Replacing COMPLETE with a word such as “Terminé” or “Completado” in eventType is not the same as displaying a localized status. The parser expects controlled tokens; keep labels in the presentation layer. Practical diagnostic: identify the source object; inspect its field path and type; compare the exact source event with the localized display; ask whether any decision-critical value or dependency changed. If it did, block release and correct the display mapping, not the source-of-truth record.

Failure 3: Conflating a dataset with a dashboard

A report card can have a friendly title while its underlying dataset keeps a different precise name. Record the card’s relation to the dataset; never make a label substitution look like a technical rename. Practical diagnostic: identify the source object; inspect its field path and type; compare the exact source event with the localized display; ask whether any decision-critical value or dependency changed. If it did, block release and correct the display mapping, not the source-of-truth record.

Failure 4: Mistranslating graph direction

A statement that orders are derived from invoices is wrong when invoices are derived from orders. Compare edge source and target pairs, not just strings. Practical diagnostic: identify the source object; inspect its field path and type; compare the exact source event with the localized display; ask whether any decision-critical value or dependency changed. If it did, block release and correct the display mapping, not the source-of-truth record.

Failure 5: Confusing job execution and nominal schedule

A midnight nominalTime does not prove the RunEvent occurred at midnight. Expose the difference in user-facing incident timelines. Practical diagnostic: identify the source object; inspect its field path and type; compare the exact source event with the localized display; ask whether any decision-critical value or dependency changed. If it did, block release and correct the display mapping, not the source-of-truth record.

Failure 6: Fusing business descriptions with physical column names

A business term such as “gross margin” may describe a calculation over several columns. A localized business glossary definition cannot replace the exact column key or its computation. Practical diagnostic: identify the source object; inspect its field path and type; compare the exact source event with the localized display; ask whether any decision-critical value or dependency changed. If it did, block release and correct the display mapping, not the source-of-truth record.

Failure 7: Suppressing indirect lineage

Filters, joins, sorting and grouping can influence results without contributing direct values. Narratives that omit these dependencies understate the pipeline’s behavior. Practical diagnostic: identify the source object; inspect its field path and type; compare the exact source event with the localized display; ask whether any decision-critical value or dependency changed. If it did, block release and correct the display mapping, not the source-of-truth record.

Failure 8: Inventing completeness

An integration that emits incomplete lineage must not be described as having a fully exhaustive graph. State extraction limits and label the source of each inferred relationship. Practical diagnostic: identify the source object; inspect its field path and type; compare the exact source event with the localized display; ask whether any decision-critical value or dependency changed. If it did, block release and correct the display mapping, not the source-of-truth record.

Failure 9: Presenting old quality measurements as current

A translated “zero missing records” message could reflect yesterday’s capture. Keep capture time, scope and metric type in the display. Practical diagnostic: identify the source object; inspect its field path and type; compare the exact source event with the localized display; ask whether any decision-critical value or dependency changed. If it did, block release and correct the display mapping, not the source-of-truth record.

Failure 10: Changing error severity by tone

FAIL and ABORT are not interchangeable and may demand different handling. A polite, simplified translation must not collapse a diagnostic distinction. Practical diagnostic: identify the source object; inspect its field path and type; compare the exact source event with the localized display; ask whether any decision-critical value or dependency changed. If it did, block release and correct the display mapping, not the source-of-truth record.

Failure 11: Confusing operational and catalog ownership

The person who emitted a lineage event, the operator of a job and the steward of the dataset may differ. Keep each role attached to the correct provenance field. Practical diagnostic: identify the source object; inspect its field path and type; compare the exact source event with the localized display; ask whether any decision-critical value or dependency changed. If it did, block release and correct the display mapping, not the source-of-truth record.

Failure 12: Blind machine translation of SQL

A translated SQL keyword or quoted identifier may turn a stored query into invalid or different code. Treat SQL as evidence; translate a separate explanation. Practical diagnostic: identify the source object; inspect its field path and type; compare the exact source event with the localized display; ask whether any decision-critical value or dependency changed. If it did, block release and correct the display mapping, not the source-of-truth record.

Failure 13: Converting time zones without labeling the conversion

Time-zone rendering is a UI calculation, not a translation of the raw timestamp. Keep canonical time and clearly label displayed local offsets. Practical diagnostic: identify the source object; inspect its field path and type; compare the exact source event with the localized display; ask whether any decision-critical value or dependency changed. If it did, block release and correct the display mapping, not the source-of-truth record.

Failure 14: Translation-memory collision

The word “source” may mean upstream input, a source-code repository, or producer provenance. Reusing one translation without context can invert the reader’s interpretation. Practical diagnostic: identify the source object; inspect its field path and type; compare the exact source event with the localized display; ask whether any decision-critical value or dependency changed. If it did, block release and correct the display mapping, not the source-of-truth record.

Failure 15: Deleting fallback language signals

If a translation is missing, showing a borrowed source-language field as if localized hides coverage defects. Mark fallback explicitly. Practical diagnostic: identify the source object; inspect its field path and type; compare the exact source event with the localized display; ask whether any decision-critical value or dependency changed. If it did, block release and correct the display mapping, not the source-of-truth record.

Failure 16: Graph node alias collision

Two datasets can share the same friendly translation. Add disambiguating technical identifiers or scope rather than merging the nodes visually. Practical diagnostic: identify the source object; inspect its field path and type; compare the exact source event with the localized display; ask whether any decision-critical value or dependency changed. If it did, block release and correct the display mapping, not the source-of-truth record.

Failure 17: Changing number representation improperly

Formatting a count with grouped digits may be acceptable for display, but rounding a precise metric or treating bytes as human-readable sizes without a declared convention can alter meaning. Practical diagnostic: identify the source object; inspect its field path and type; compare the exact source event with the localized display; ask whether any decision-critical value or dependency changed. If it did, block release and correct the display mapping, not the source-of-truth record.

Failure 18: Ignoring facet schema versions

A similarly named facet can acquire fields or different semantics over releases. Translate against the actual referenced schema rather than memorized examples. Practical diagnostic: identify the source object; inspect its field path and type; compare the exact source event with the localized display; ask whether any decision-critical value or dependency changed. If it did, block release and correct the display mapping, not the source-of-truth record.

Failure 19: Equating an assertion with a measurement

A test pass/fail report and a numeric nullCount tell different stories. Preserve the difference between measured quantities and evaluated rules. Practical diagnostic: identify the source object; inspect its field path and type; compare the exact source event with the localized display; ask whether any decision-critical value or dependency changed. If it did, block release and correct the display mapping, not the source-of-truth record.

Failure 20: Treating a custom facet as standard

A vendor-defined facet can use local terms or experimental semantics. Require the producer’s schema and owner approval rather than assigning an official OpenLineage definition. Practical diagnostic: identify the source object; inspect its field path and type; compare the exact source event with the localized display; ask whether any decision-critical value or dependency changed. If it did, block release and correct the display mapping, not the source-of-truth record.

Failure 21: Exposing secrets through the translation workflow

Lineage metadata may contain sensitive resource names, SQL fragments or internal paths. Apply access controls and redact safely for translation without silently changing the underlying evidence. Practical diagnostic: identify the source object; inspect its field path and type; compare the exact source event with the localized display; ask whether any decision-critical value or dependency changed. If it did, block release and correct the display mapping, not the source-of-truth record.

Failure 22: Using a translation to conceal a migration

If a dataset truly changes identity, the engineering team must migrate and document it. A localization-only change must not conceal a topology modification. Practical diagnostic: identify the source object; inspect its field path and type; compare the exact source event with the localized display; ask whether any decision-critical value or dependency changed. If it did, block release and correct the display mapping, not the source-of-truth record.

Failure 23: Assuming locale-neutral abbreviations

Acronyms and short codes can have different expansions across audiences. Define each in context once; do not translate code tokens themselves. Practical diagnostic: identify the source object; inspect its field path and type; compare the exact source event with the localized display; ask whether any decision-critical value or dependency changed. If it did, block release and correct the display mapping, not the source-of-truth record.

Failure 24: No reproducible acceptance test

A review signed off from screenshots alone may miss a graph-edge mismatch. Verify the same event fixtures and deterministic entity IDs before and after locale switching. Practical diagnostic: identify the source object; inspect its field path and type; compare the exact source event with the localized display; ask whether any decision-critical value or dependency changed. If it did, block release and correct the display mapping, not the source-of-truth record.

Worked diagnostics: thirty technical modules, six reader situations each

The following 180 fictional diagnostics are constructed as teaching fixtures. Dataset names, event IDs, example labels, incidents and reviewers are invented. They are not live customer records and they do not demonstrate an implementation’s actual behavior. Each test moves through six distinct situations: catalog display, machine translation, conflicting evidence, incident response, bilingual review and release regression. The objective is to practise a repeatable reasoning method on genuinely different parts of an OpenLineage record.

Identity, execution and event provenance

Module 01: Run state enumeration

In a nightly invoice reconciliation job, the source fixture is eventType = COMPLETE. The tempting mistranslation is eventType = CONCLUÍDO, whereas the controlled explanation is “Completed” as a localized display label while the event value stays COMPLETE. This distinction matters because eventType is one of six protocol-defined state tokens, not ordinary narrative prose. A person who does not know the domain can easily overlook that OTHER can carry supplementary metadata without marking a terminal result. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 001: Catalog label versus identifier

A reader opens a catalog card for a nightly invoice reconciliation job. The underlying evidence is eventType = COMPLETE. The dangerous shortcut is eventType = CONCLUÍDO, because it treats a localized display convention as if it were the object itself. A better presentation is “Completed” as a localized display label while the event value stays COMPLETE. The primary contract is: eventType is one of six protocol-defined state tokens, not ordinary narrative prose. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, compare the START and terminal events for the same runId. The language editor should also note that OTHER can carry supplementary metadata without marking a terminal result. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 002: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a nightly invoice reconciliation job as ordinary text. It may output eventType = CONCLUÍDO, a grammatically plausible change that does not respect the protocol role of eventType = COMPLETE. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Completed” as a localized display label while the event value stays COMPLETE. For this field the reason is not stylistic preference: eventType is one of six protocol-defined state tokens, not ordinary narrative prose. A reviewer should apply the concrete test to this example: compare the START and terminal events for the same runId. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that OTHER can carry supplementary metadata without marking a terminal result.

Diagnostic 003: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a nightly invoice reconciliation job; the resulting translation resembles eventType = CONCLUÍDO. Reconstruct the full path and the scenario, a nightly invoice reconciliation job, before rewriting the label. Put the original evidence, eventType = COMPLETE, next to the human recommendation, “Completed” as a localized display label while the event value stays COMPLETE. Then ask whether the source record actually supports the claim: eventType is one of six protocol-defined state tokens, not ordinary narrative prose. A correct diagnosis is reproducible because compare the START and terminal events for the same runId. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: OTHER can carry supplementary metadata without marking a terminal result. The goal is reader understanding without inventing a technical fact.

Diagnostic 004: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a nightly invoice reconciliation job. If the translation has followed the shortcut eventType = CONCLUÍDO, the operational danger is that a support analyst may confuse a finished run with an aborted run. A better answer begins with the source evidence eventType = COMPLETE, describes it as “Completed” as a localized display label while the event value stays COMPLETE, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: compare the START and terminal events for the same runId. The caveat OTHER can carry supplementary metadata without marking a terminal result also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 005: Bilingual reviewer handoff

Give an independent domain reviewer the source item eventType = COMPLETE, the rejected wording eventType = CONCLUÍDO, and the proposed user-facing wording “Completed” as a localized display label while the event value stays COMPLETE. Ask the reviewer to explain in plain language why eventType is one of six protocol-defined state tokens, not ordinary narrative prose. They should supply an example from a nightly invoice reconciliation job, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: compare the START and terminal events for the same runId. A final note should state OTHER can carry supplementary metadata without marking a terminal result, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 006: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a nightly invoice reconciliation job. The technical reference eventType = COMPLETE must still identify the same object or evidence, while the localized display may read “Completed” as a localized display label while the event value stays COMPLETE. Fail the test if the export instead contains eventType = CONCLUÍDO where an identifier or protocol value was expected. Run the field-specific assertion: compare the START and terminal events for the same runId. This matters because a support analyst may confuse a finished run with an aborted run. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation OTHER can carry supplementary metadata without marking a terminal result into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify run state enumeration, but it must retain the source evidence and must pass this field-specific test: compare the START and terminal events for the same runId. Transfer destination: status badges in an incident console.

Module 02: Run UUID

In a restarted customer-load task, the source fixture is run.runId = 7aa0-25d9-4f76-a309. The tempting mistranslation is run.runId = translated or newly generated UUID, whereas the controlled explanation is “Customer import, attempt two” as an alias keyed to the unchanged original UUID. This distinction matters because the run identifier correlates a single execution’s events; it is not a title. A person who does not know the domain can easily overlook that a job can have many runs but a run belongs to the identity of that execution. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 007: Catalog label versus identifier

A reader opens a catalog card for a restarted customer-load task. The underlying evidence is run.runId = 7aa0-25d9-4f76-a309. The dangerous shortcut is run.runId = translated or newly generated UUID, because it treats a localized display convention as if it were the object itself. A better presentation is “Customer import, attempt two” as an alias keyed to the unchanged original UUID. The primary contract is: the run identifier correlates a single execution’s events; it is not a title. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, join START and FAIL records using the untouched runId in the original fixture. The language editor should also note that a job can have many runs but a run belongs to the identity of that execution. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 008: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a restarted customer-load task as ordinary text. It may output run.runId = translated or newly generated UUID, a grammatically plausible change that does not respect the protocol role of run.runId = 7aa0-25d9-4f76-a309. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Customer import, attempt two” as an alias keyed to the unchanged original UUID. For this field the reason is not stylistic preference: the run identifier correlates a single execution’s events; it is not a title. A reviewer should apply the concrete test to this example: join START and FAIL records using the untouched runId in the original fixture. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that a job can have many runs but a run belongs to the identity of that execution.

Diagnostic 009: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a restarted customer-load task; the resulting translation resembles run.runId = translated or newly generated UUID. Reconstruct the full path and the scenario, a restarted customer-load task, before rewriting the label. Put the original evidence, run.runId = 7aa0-25d9-4f76-a309, next to the human recommendation, “Customer import, attempt two” as an alias keyed to the unchanged original UUID. Then ask whether the source record actually supports the claim: the run identifier correlates a single execution’s events; it is not a title. A correct diagnosis is reproducible because join START and FAIL records using the untouched runId in the original fixture. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: a job can have many runs but a run belongs to the identity of that execution. The goal is reader understanding without inventing a technical fact.

Diagnostic 010: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a restarted customer-load task. If the translation has followed the shortcut run.runId = translated or newly generated UUID, the operational danger is that a retry could be mistaken for the first failed execution. A better answer begins with the source evidence run.runId = 7aa0-25d9-4f76-a309, describes it as “Customer import, attempt two” as an alias keyed to the unchanged original UUID, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: join START and FAIL records using the untouched runId in the original fixture. The caveat a job can have many runs but a run belongs to the identity of that execution also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 011: Bilingual reviewer handoff

Give an independent domain reviewer the source item run.runId = 7aa0-25d9-4f76-a309, the rejected wording run.runId = translated or newly generated UUID, and the proposed user-facing wording “Customer import, attempt two” as an alias keyed to the unchanged original UUID. Ask the reviewer to explain in plain language why the run identifier correlates a single execution’s events; it is not a title. They should supply an example from a restarted customer-load task, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: join START and FAIL records using the untouched runId in the original fixture. A final note should state a job can have many runs but a run belongs to the identity of that execution, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 012: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a restarted customer-load task. The technical reference run.runId = 7aa0-25d9-4f76-a309 must still identify the same object or evidence, while the localized display may read “Customer import, attempt two” as an alias keyed to the unchanged original UUID. Fail the test if the export instead contains run.runId = translated or newly generated UUID where an identifier or protocol value was expected. Run the field-specific assertion: join START and FAIL records using the untouched runId in the original fixture. This matters because a retry could be mistaken for the first failed execution. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation a job can have many runs but a run belongs to the identity of that execution into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify run uuid, but it must retain the source evidence and must pass this field-specific test: join START and FAIL records using the untouched runId in the original fixture. Transfer destination: a timeline that combines batch retries.

Module 03: Event timestamp

In a delayed inventory aggregation, the source fixture is eventTime = 2026-10-04T05:15:20Z. The tempting mistranslation is “scheduled for 05:15” presented as observed event time, whereas the controlled explanation is “Event received at 05:15:20 UTC” with the canonical string retained. This distinction matters because eventTime is an observation timestamp, not automatically the business date. A person who does not know the domain can easily overlook that local time-zone conversion must display its offset and cannot erase UTC evidence. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 013: Catalog label versus identifier

A reader opens a catalog card for a delayed inventory aggregation. The underlying evidence is eventTime = 2026-10-04T05:15:20Z. The dangerous shortcut is “scheduled for 05:15” presented as observed event time, because it treats a localized display convention as if it were the object itself. A better presentation is “Event received at 05:15:20 UTC” with the canonical string retained. The primary contract is: eventTime is an observation timestamp, not automatically the business date. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, compare raw UTC, displayed zone and log collection timing. The language editor should also note that local time-zone conversion must display its offset and cannot erase UTC evidence. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 014: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a delayed inventory aggregation as ordinary text. It may output “scheduled for 05:15” presented as observed event time, a grammatically plausible change that does not respect the protocol role of eventTime = 2026-10-04T05:15:20Z. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Event received at 05:15:20 UTC” with the canonical string retained. For this field the reason is not stylistic preference: eventTime is an observation timestamp, not automatically the business date. A reviewer should apply the concrete test to this example: compare raw UTC, displayed zone and log collection timing. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that local time-zone conversion must display its offset and cannot erase UTC evidence.

Diagnostic 015: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a delayed inventory aggregation; the resulting translation resembles “scheduled for 05:15” presented as observed event time. Reconstruct the full path and the scenario, a delayed inventory aggregation, before rewriting the label. Put the original evidence, eventTime = 2026-10-04T05:15:20Z, next to the human recommendation, “Event received at 05:15:20 UTC” with the canonical string retained. Then ask whether the source record actually supports the claim: eventTime is an observation timestamp, not automatically the business date. A correct diagnosis is reproducible because compare raw UTC, displayed zone and log collection timing. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: local time-zone conversion must display its offset and cannot erase UTC evidence. The goal is reader understanding without inventing a technical fact.

Diagnostic 016: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a delayed inventory aggregation. If the translation has followed the shortcut “scheduled for 05:15” presented as observed event time, the operational danger is that an analyst may infer the job ran on schedule when it started late. A better answer begins with the source evidence eventTime = 2026-10-04T05:15:20Z, describes it as “Event received at 05:15:20 UTC” with the canonical string retained, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: compare raw UTC, displayed zone and log collection timing. The caveat local time-zone conversion must display its offset and cannot erase UTC evidence also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 017: Bilingual reviewer handoff

Give an independent domain reviewer the source item eventTime = 2026-10-04T05:15:20Z, the rejected wording “scheduled for 05:15” presented as observed event time, and the proposed user-facing wording “Event received at 05:15:20 UTC” with the canonical string retained. Ask the reviewer to explain in plain language why eventTime is an observation timestamp, not automatically the business date. They should supply an example from a delayed inventory aggregation, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: compare raw UTC, displayed zone and log collection timing. A final note should state local time-zone conversion must display its offset and cannot erase UTC evidence, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 018: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a delayed inventory aggregation. The technical reference eventTime = 2026-10-04T05:15:20Z must still identify the same object or evidence, while the localized display may read “Event received at 05:15:20 UTC” with the canonical string retained. Fail the test if the export instead contains “scheduled for 05:15” presented as observed event time where an identifier or protocol value was expected. Run the field-specific assertion: compare raw UTC, displayed zone and log collection timing. This matters because an analyst may infer the job ran on schedule when it started late. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation local time-zone conversion must display its offset and cannot erase UTC evidence into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify event timestamp, but it must retain the source evidence and must pass this field-specific test: compare raw UTC, displayed zone and log collection timing. Transfer destination: a multilingual incident chronology.

Module 04: Event producer

In a Python lineage emitter attached to an ETL runner, the source fixture is producer = https://example.org/emitters/python/2.4. The tempting mistranslation is producer displayed as the data owner, whereas the controlled explanation is “Metadata emitted by Python connector 2.4” in explanatory text. This distinction matters because the producer URI identifies the metadata-producing software, not the dataset steward. A person who does not know the domain can easily overlook that a producer can emit information about systems it does not operate. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 019: Catalog label versus identifier

A reader opens a catalog card for a Python lineage emitter attached to an ETL runner. The underlying evidence is producer = https://example.org/emitters/python/2.4. The dangerous shortcut is producer displayed as the data owner, because it treats a localized display convention as if it were the object itself. A better presentation is “Metadata emitted by Python connector 2.4” in explanatory text. The primary contract is: the producer URI identifies the metadata-producing software, not the dataset steward. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, resolve the producer record and compare the ownership facet separately. The language editor should also note that a producer can emit information about systems it does not operate. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 020: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a Python lineage emitter attached to an ETL runner as ordinary text. It may output producer displayed as the data owner, a grammatically plausible change that does not respect the protocol role of producer = https://example.org/emitters/python/2.4. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Metadata emitted by Python connector 2.4” in explanatory text. For this field the reason is not stylistic preference: the producer URI identifies the metadata-producing software, not the dataset steward. A reviewer should apply the concrete test to this example: resolve the producer record and compare the ownership facet separately. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that a producer can emit information about systems it does not operate.

Diagnostic 021: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a Python lineage emitter attached to an ETL runner; the resulting translation resembles producer displayed as the data owner. Reconstruct the full path and the scenario, a Python lineage emitter attached to an ETL runner, before rewriting the label. Put the original evidence, producer = https://example.org/emitters/python/2.4, next to the human recommendation, “Metadata emitted by Python connector 2.4” in explanatory text. Then ask whether the source record actually supports the claim: the producer URI identifies the metadata-producing software, not the dataset steward. A correct diagnosis is reproducible because resolve the producer record and compare the ownership facet separately. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: a producer can emit information about systems it does not operate. The goal is reader understanding without inventing a technical fact.

Diagnostic 022: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a Python lineage emitter attached to an ETL runner. If the translation has followed the shortcut producer displayed as the data owner, the operational danger is that the wrong team may be blamed for a malformed lineage report. A better answer begins with the source evidence producer = https://example.org/emitters/python/2.4, describes it as “Metadata emitted by Python connector 2.4” in explanatory text, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: resolve the producer record and compare the ownership facet separately. The caveat a producer can emit information about systems it does not operate also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 023: Bilingual reviewer handoff

Give an independent domain reviewer the source item producer = https://example.org/emitters/python/2.4, the rejected wording producer displayed as the data owner, and the proposed user-facing wording “Metadata emitted by Python connector 2.4” in explanatory text. Ask the reviewer to explain in plain language why the producer URI identifies the metadata-producing software, not the dataset steward. They should supply an example from a Python lineage emitter attached to an ETL runner, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: resolve the producer record and compare the ownership facet separately. A final note should state a producer can emit information about systems it does not operate, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 024: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a Python lineage emitter attached to an ETL runner. The technical reference producer = https://example.org/emitters/python/2.4 must still identify the same object or evidence, while the localized display may read “Metadata emitted by Python connector 2.4” in explanatory text. Fail the test if the export instead contains producer displayed as the data owner where an identifier or protocol value was expected. Run the field-specific assertion: resolve the producer record and compare the ownership facet separately. This matters because the wrong team may be blamed for a malformed lineage report. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation a producer can emit information about systems it does not operate into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify event producer, but it must retain the source evidence and must pass this field-specific test: resolve the producer record and compare the ownership facet separately. Transfer destination: a technical support escalation.

Module 05: Event schema pointer

In a client moving from one event schema revision to another, the source fixture is schemaURL = https://openlineage.io/spec/2-0-2/OpenLineage.json#/$defs/RunEvent. The tempting mistranslation is schemaURL converted into a human-facing page slug, whereas the controlled explanation is “Event schema revision used for validation” as a caption only. This distinction matters because a versioned schema URL is a machine reference; translation cannot change its address. A person who does not know the domain can easily overlook that a documentation page can change while the event still points to a particular schema. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 025: Catalog label versus identifier

A reader opens a catalog card for a client moving from one event schema revision to another. The underlying evidence is schemaURL = https://openlineage.io/spec/2-0-2/OpenLineage.json#/$defs/RunEvent. The dangerous shortcut is schemaURL converted into a human-facing page slug, because it treats a localized display convention as if it were the object itself. A better presentation is “Event schema revision used for validation” as a caption only. The primary contract is: a versioned schema URL is a machine reference; translation cannot change its address. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, validate the exact event against the referenced schema and inspect the URL bytes. The language editor should also note that a documentation page can change while the event still points to a particular schema. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 026: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a client moving from one event schema revision to another as ordinary text. It may output schemaURL converted into a human-facing page slug, a grammatically plausible change that does not respect the protocol role of schemaURL = https://openlineage.io/spec/2-0-2/OpenLineage.json#/$defs/RunEvent. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Event schema revision used for validation” as a caption only. For this field the reason is not stylistic preference: a versioned schema URL is a machine reference; translation cannot change its address. A reviewer should apply the concrete test to this example: validate the exact event against the referenced schema and inspect the URL bytes. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that a documentation page can change while the event still points to a particular schema.

Diagnostic 027: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a client moving from one event schema revision to another; the resulting translation resembles schemaURL converted into a human-facing page slug. Reconstruct the full path and the scenario, a client moving from one event schema revision to another, before rewriting the label. Put the original evidence, schemaURL = https://openlineage.io/spec/2-0-2/OpenLineage.json#/$defs/RunEvent, next to the human recommendation, “Event schema revision used for validation” as a caption only. Then ask whether the source record actually supports the claim: a versioned schema URL is a machine reference; translation cannot change its address. A correct diagnosis is reproducible because validate the exact event against the referenced schema and inspect the URL bytes. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: a documentation page can change while the event still points to a particular schema. The goal is reader understanding without inventing a technical fact.

Diagnostic 028: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a client moving from one event schema revision to another. If the translation has followed the shortcut schemaURL converted into a human-facing page slug, the operational danger is that a consuming service may reject the whole event. A better answer begins with the source evidence schemaURL = https://openlineage.io/spec/2-0-2/OpenLineage.json#/$defs/RunEvent, describes it as “Event schema revision used for validation” as a caption only, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: validate the exact event against the referenced schema and inspect the URL bytes. The caveat a documentation page can change while the event still points to a particular schema also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 029: Bilingual reviewer handoff

Give an independent domain reviewer the source item schemaURL = https://openlineage.io/spec/2-0-2/OpenLineage.json#/$defs/RunEvent, the rejected wording schemaURL converted into a human-facing page slug, and the proposed user-facing wording “Event schema revision used for validation” as a caption only. Ask the reviewer to explain in plain language why a versioned schema URL is a machine reference; translation cannot change its address. They should supply an example from a client moving from one event schema revision to another, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: validate the exact event against the referenced schema and inspect the URL bytes. A final note should state a documentation page can change while the event still points to a particular schema, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 030: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a client moving from one event schema revision to another. The technical reference schemaURL = https://openlineage.io/spec/2-0-2/OpenLineage.json#/$defs/RunEvent must still identify the same object or evidence, while the localized display may read “Event schema revision used for validation” as a caption only. Fail the test if the export instead contains schemaURL converted into a human-facing page slug where an identifier or protocol value was expected. Run the field-specific assertion: validate the exact event against the referenced schema and inspect the URL bytes. This matters because a consuming service may reject the whole event. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation a documentation page can change while the event still points to a particular schema into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify event schema pointer, but it must retain the source evidence and must pass this field-specific test: validate the exact event against the referenced schema and inspect the URL bytes. Transfer destination: a compatibility release checklist.

Module 06: Job namespace

In an orchestrated billing pipeline, the source fixture is job.namespace = workflows-production. The tempting mistranslation is job.namespace = flux-de-travail-production, whereas the controlled explanation is “Production workflows” as a display alias with namespace untouched. This distinction matters because namespace plus name identifies a job in the model. A person who does not know the domain can easily overlook that same job names may exist safely in distinct namespaces. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 031: Catalog label versus identifier

A reader opens a catalog card for an orchestrated billing pipeline. The underlying evidence is job.namespace = workflows-production. The dangerous shortcut is job.namespace = flux-de-travail-production, because it treats a localized display convention as if it were the object itself. A better presentation is “Production workflows” as a display alias with namespace untouched. The primary contract is: namespace plus name identifies a job in the model. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, compare identity keys in event snapshots before and after locale change. The language editor should also note that same job names may exist safely in distinct namespaces. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 032: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for an orchestrated billing pipeline as ordinary text. It may output job.namespace = flux-de-travail-production, a grammatically plausible change that does not respect the protocol role of job.namespace = workflows-production. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Production workflows” as a display alias with namespace untouched. For this field the reason is not stylistic preference: namespace plus name identifies a job in the model. A reviewer should apply the concrete test to this example: compare identity keys in event snapshots before and after locale change. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that same job names may exist safely in distinct namespaces.

Diagnostic 033: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of an orchestrated billing pipeline; the resulting translation resembles job.namespace = flux-de-travail-production. Reconstruct the full path and the scenario, an orchestrated billing pipeline, before rewriting the label. Put the original evidence, job.namespace = workflows-production, next to the human recommendation, “Production workflows” as a display alias with namespace untouched. Then ask whether the source record actually supports the claim: namespace plus name identifies a job in the model. A correct diagnosis is reproducible because compare identity keys in event snapshots before and after locale change. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: same job names may exist safely in distinct namespaces. The goal is reader understanding without inventing a technical fact.

Diagnostic 034: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in an orchestrated billing pipeline. If the translation has followed the shortcut job.namespace = flux-de-travail-production, the operational danger is that graph nodes may split into two jobs. A better answer begins with the source evidence job.namespace = workflows-production, describes it as “Production workflows” as a display alias with namespace untouched, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: compare identity keys in event snapshots before and after locale change. The caveat same job names may exist safely in distinct namespaces also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 035: Bilingual reviewer handoff

Give an independent domain reviewer the source item job.namespace = workflows-production, the rejected wording job.namespace = flux-de-travail-production, and the proposed user-facing wording “Production workflows” as a display alias with namespace untouched. Ask the reviewer to explain in plain language why namespace plus name identifies a job in the model. They should supply an example from an orchestrated billing pipeline, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: compare identity keys in event snapshots before and after locale change. A final note should state same job names may exist safely in distinct namespaces, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 036: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for an orchestrated billing pipeline. The technical reference job.namespace = workflows-production must still identify the same object or evidence, while the localized display may read “Production workflows” as a display alias with namespace untouched. Fail the test if the export instead contains job.namespace = flux-de-travail-production where an identifier or protocol value was expected. Run the field-specific assertion: compare identity keys in event snapshots before and after locale change. This matters because graph nodes may split into two jobs. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation same job names may exist safely in distinct namespaces into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify job namespace, but it must retain the source evidence and must pass this field-specific test: compare identity keys in event snapshots before and after locale change. Transfer destination: a multilingual job search index.

Module 07: Job technical name

In a daily revenue transform, the source fixture is job.name = compute_daily_revenue. The tempting mistranslation is job.name = calculer_recette_quotidienne, whereas the controlled explanation is “Calculate daily revenue” as a human-readable job title. This distinction matters because a technical job name may be referenced by orchestration and graph joins. A person who does not know the domain can easily overlook that friendly names can change more frequently than code identities. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 037: Catalog label versus identifier

A reader opens a catalog card for a daily revenue transform. The underlying evidence is job.name = compute_daily_revenue. The dangerous shortcut is job.name = calculer_recette_quotidienne, because it treats a localized display convention as if it were the object itself. A better presentation is “Calculate daily revenue” as a human-readable job title. The primary contract is: a technical job name may be referenced by orchestration and graph joins. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, retrieve a run by exact namespace/name and check its outputs. The language editor should also note that friendly names can change more frequently than code identities. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 038: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a daily revenue transform as ordinary text. It may output job.name = calculer_recette_quotidienne, a grammatically plausible change that does not respect the protocol role of job.name = compute_daily_revenue. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Calculate daily revenue” as a human-readable job title. For this field the reason is not stylistic preference: a technical job name may be referenced by orchestration and graph joins. A reviewer should apply the concrete test to this example: retrieve a run by exact namespace/name and check its outputs. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that friendly names can change more frequently than code identities.

Diagnostic 039: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a daily revenue transform; the resulting translation resembles job.name = calculer_recette_quotidienne. Reconstruct the full path and the scenario, a daily revenue transform, before rewriting the label. Put the original evidence, job.name = compute_daily_revenue, next to the human recommendation, “Calculate daily revenue” as a human-readable job title. Then ask whether the source record actually supports the claim: a technical job name may be referenced by orchestration and graph joins. A correct diagnosis is reproducible because retrieve a run by exact namespace/name and check its outputs. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: friendly names can change more frequently than code identities. The goal is reader understanding without inventing a technical fact.

Diagnostic 040: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a daily revenue transform. If the translation has followed the shortcut job.name = calculer_recette_quotidienne, the operational danger is that the localized title may be mistaken for a newly deployed task. A better answer begins with the source evidence job.name = compute_daily_revenue, describes it as “Calculate daily revenue” as a human-readable job title, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: retrieve a run by exact namespace/name and check its outputs. The caveat friendly names can change more frequently than code identities also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 041: Bilingual reviewer handoff

Give an independent domain reviewer the source item job.name = compute_daily_revenue, the rejected wording job.name = calculer_recette_quotidienne, and the proposed user-facing wording “Calculate daily revenue” as a human-readable job title. Ask the reviewer to explain in plain language why a technical job name may be referenced by orchestration and graph joins. They should supply an example from a daily revenue transform, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: retrieve a run by exact namespace/name and check its outputs. A final note should state friendly names can change more frequently than code identities, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 042: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a daily revenue transform. The technical reference job.name = compute_daily_revenue must still identify the same object or evidence, while the localized display may read “Calculate daily revenue” as a human-readable job title. Fail the test if the export instead contains job.name = calculer_recette_quotidienne where an identifier or protocol value was expected. Run the field-specific assertion: retrieve a run by exact namespace/name and check its outputs. This matters because the localized title may be mistaken for a newly deployed task. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation friendly names can change more frequently than code identities into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify job technical name, but it must retain the source evidence and must pass this field-specific test: retrieve a run by exact namespace/name and check its outputs. Transfer destination: an operations dashboard and a runbook.

Module 08: Input Dataset namespace

In a database receiving raw shopping-cart events, the source fixture is inputs[0].namespace = postgresql://db.example.org:5432. The tempting mistranslation is the database URI translated into prose in structured JSON, whereas the controlled explanation is “Primary transactional database” in a separate description. This distinction matters because namespace establishes the scope of a Dataset identity. A person who does not know the domain can easily overlook that different environments can carry identical table names. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 043: Catalog label versus identifier

A reader opens a catalog card for a database receiving raw shopping-cart events. The underlying evidence is inputs[0].namespace = postgresql://db.example.org:5432. The dangerous shortcut is the database URI translated into prose in structured JSON, because it treats a localized display convention as if it were the object itself. A better presentation is “Primary transactional database” in a separate description. The primary contract is: namespace establishes the scope of a Dataset identity. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, compare all input namespace URIs in the original and exported localized record. The language editor should also note that different environments can carry identical table names. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 044: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a database receiving raw shopping-cart events as ordinary text. It may output the database URI translated into prose in structured JSON, a grammatically plausible change that does not respect the protocol role of inputs[0].namespace = postgresql://db.example.org:5432. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Primary transactional database” in a separate description. For this field the reason is not stylistic preference: namespace establishes the scope of a Dataset identity. A reviewer should apply the concrete test to this example: compare all input namespace URIs in the original and exported localized record. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that different environments can carry identical table names.

Diagnostic 045: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a database receiving raw shopping-cart events; the resulting translation resembles the database URI translated into prose in structured JSON. Reconstruct the full path and the scenario, a database receiving raw shopping-cart events, before rewriting the label. Put the original evidence, inputs[0].namespace = postgresql://db.example.org:5432, next to the human recommendation, “Primary transactional database” in a separate description. Then ask whether the source record actually supports the claim: namespace establishes the scope of a Dataset identity. A correct diagnosis is reproducible because compare all input namespace URIs in the original and exported localized record. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: different environments can carry identical table names. The goal is reader understanding without inventing a technical fact.

Diagnostic 046: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a database receiving raw shopping-cart events. If the translation has followed the shortcut the database URI translated into prose in structured JSON, the operational danger is that two database instances may be merged visually or separated falsely. A better answer begins with the source evidence inputs[0].namespace = postgresql://db.example.org:5432, describes it as “Primary transactional database” in a separate description, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: compare all input namespace URIs in the original and exported localized record. The caveat different environments can carry identical table names also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 047: Bilingual reviewer handoff

Give an independent domain reviewer the source item inputs[0].namespace = postgresql://db.example.org:5432, the rejected wording the database URI translated into prose in structured JSON, and the proposed user-facing wording “Primary transactional database” in a separate description. Ask the reviewer to explain in plain language why namespace establishes the scope of a Dataset identity. They should supply an example from a database receiving raw shopping-cart events, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: compare all input namespace URIs in the original and exported localized record. A final note should state different environments can carry identical table names, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 048: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a database receiving raw shopping-cart events. The technical reference inputs[0].namespace = postgresql://db.example.org:5432 must still identify the same object or evidence, while the localized display may read “Primary transactional database” in a separate description. Fail the test if the export instead contains the database URI translated into prose in structured JSON where an identifier or protocol value was expected. Run the field-specific assertion: compare all input namespace URIs in the original and exported localized record. This matters because two database instances may be merged visually or separated falsely. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation different environments can carry identical table names into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify input dataset namespace, but it must retain the source evidence and must pass this field-specific test: compare all input namespace URIs in the original and exported localized record. Transfer destination: a data catalog linking test and production datasets.

Module 09: Input Dataset name

In a purchases feed consumed by a transformation, the source fixture is inputs[0].name = raw.orders. The tempting mistranslation is inputs[0].name = bruto.pedidos, whereas the controlled explanation is “Raw orders” as a reader label mapped to raw.orders. This distinction matters because input name is part of an exact dataset reference. A person who does not know the domain can easily overlook that a translated label must not imply that a view and table are interchangeable. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 049: Catalog label versus identifier

A reader opens a catalog card for a purchases feed consumed by a transformation. The underlying evidence is inputs[0].name = raw.orders. The dangerous shortcut is inputs[0].name = bruto.pedidos, because it treats a localized display convention as if it were the object itself. A better presentation is “Raw orders” as a reader label mapped to raw.orders. The primary contract is: input name is part of an exact dataset reference. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, match the source dataset node ID and its original source-system reference. The language editor should also note that a translated label must not imply that a view and table are interchangeable. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 050: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a purchases feed consumed by a transformation as ordinary text. It may output inputs[0].name = bruto.pedidos, a grammatically plausible change that does not respect the protocol role of inputs[0].name = raw.orders. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Raw orders” as a reader label mapped to raw.orders. For this field the reason is not stylistic preference: input name is part of an exact dataset reference. A reviewer should apply the concrete test to this example: match the source dataset node ID and its original source-system reference. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that a translated label must not imply that a view and table are interchangeable.

Diagnostic 051: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a purchases feed consumed by a transformation; the resulting translation resembles inputs[0].name = bruto.pedidos. Reconstruct the full path and the scenario, a purchases feed consumed by a transformation, before rewriting the label. Put the original evidence, inputs[0].name = raw.orders, next to the human recommendation, “Raw orders” as a reader label mapped to raw.orders. Then ask whether the source record actually supports the claim: input name is part of an exact dataset reference. A correct diagnosis is reproducible because match the source dataset node ID and its original source-system reference. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: a translated label must not imply that a view and table are interchangeable. The goal is reader understanding without inventing a technical fact.

Diagnostic 052: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a purchases feed consumed by a transformation. If the translation has followed the shortcut inputs[0].name = bruto.pedidos, the operational danger is that a user may investigate the wrong upstream table. A better answer begins with the source evidence inputs[0].name = raw.orders, describes it as “Raw orders” as a reader label mapped to raw.orders, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: match the source dataset node ID and its original source-system reference. The caveat a translated label must not imply that a view and table are interchangeable also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 053: Bilingual reviewer handoff

Give an independent domain reviewer the source item inputs[0].name = raw.orders, the rejected wording inputs[0].name = bruto.pedidos, and the proposed user-facing wording “Raw orders” as a reader label mapped to raw.orders. Ask the reviewer to explain in plain language why input name is part of an exact dataset reference. They should supply an example from a purchases feed consumed by a transformation, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: match the source dataset node ID and its original source-system reference. A final note should state a translated label must not imply that a view and table are interchangeable, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 054: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a purchases feed consumed by a transformation. The technical reference inputs[0].name = raw.orders must still identify the same object or evidence, while the localized display may read “Raw orders” as a reader label mapped to raw.orders. Fail the test if the export instead contains inputs[0].name = bruto.pedidos where an identifier or protocol value was expected. Run the field-specific assertion: match the source dataset node ID and its original source-system reference. This matters because a user may investigate the wrong upstream table. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation a translated label must not imply that a view and table are interchangeable into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify input dataset name, but it must retain the source evidence and must pass this field-specific test: match the source dataset node ID and its original source-system reference. Transfer destination: upstream lineage navigation.

Module 10: Output Dataset name

In a completed customer cohort calculation, the source fixture is outputs[0].name = analytics.customer_cohorts. The tempting mistranslation is outputs[0].name = analyse.cohortes_clients, whereas the controlled explanation is “Customer cohorts” as a display title with explicit technical key. This distinction matters because an output dataset is a produced entity, not a label to overwrite. A person who does not know the domain can easily overlook that output labels can be pleasant without promising the metric was validated. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 055: Catalog label versus identifier

A reader opens a catalog card for a completed customer cohort calculation. The underlying evidence is outputs[0].name = analytics.customer_cohorts. The dangerous shortcut is outputs[0].name = analyse.cohortes_clients, because it treats a localized display convention as if it were the object itself. A better presentation is “Customer cohorts” as a display title with explicit technical key. The primary contract is: an output dataset is a produced entity, not a label to overwrite. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, compare resulting edge list from original versus target locale. The language editor should also note that output labels can be pleasant without promising the metric was validated. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 056: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a completed customer cohort calculation as ordinary text. It may output outputs[0].name = analyse.cohortes_clients, a grammatically plausible change that does not respect the protocol role of outputs[0].name = analytics.customer_cohorts. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Customer cohorts” as a display title with explicit technical key. For this field the reason is not stylistic preference: an output dataset is a produced entity, not a label to overwrite. A reviewer should apply the concrete test to this example: compare resulting edge list from original versus target locale. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that output labels can be pleasant without promising the metric was validated.

Diagnostic 057: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a completed customer cohort calculation; the resulting translation resembles outputs[0].name = analyse.cohortes_clients. Reconstruct the full path and the scenario, a completed customer cohort calculation, before rewriting the label. Put the original evidence, outputs[0].name = analytics.customer_cohorts, next to the human recommendation, “Customer cohorts” as a display title with explicit technical key. Then ask whether the source record actually supports the claim: an output dataset is a produced entity, not a label to overwrite. A correct diagnosis is reproducible because compare resulting edge list from original versus target locale. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: output labels can be pleasant without promising the metric was validated. The goal is reader understanding without inventing a technical fact.

Diagnostic 058: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a completed customer cohort calculation. If the translation has followed the shortcut outputs[0].name = analyse.cohortes_clients, the operational danger is that business teams may confuse new cohorts with source customer data. A better answer begins with the source evidence outputs[0].name = analytics.customer_cohorts, describes it as “Customer cohorts” as a display title with explicit technical key, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: compare resulting edge list from original versus target locale. The caveat output labels can be pleasant without promising the metric was validated also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 059: Bilingual reviewer handoff

Give an independent domain reviewer the source item outputs[0].name = analytics.customer_cohorts, the rejected wording outputs[0].name = analyse.cohortes_clients, and the proposed user-facing wording “Customer cohorts” as a display title with explicit technical key. Ask the reviewer to explain in plain language why an output dataset is a produced entity, not a label to overwrite. They should supply an example from a completed customer cohort calculation, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: compare resulting edge list from original versus target locale. A final note should state output labels can be pleasant without promising the metric was validated, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 060: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a completed customer cohort calculation. The technical reference outputs[0].name = analytics.customer_cohorts must still identify the same object or evidence, while the localized display may read “Customer cohorts” as a display title with explicit technical key. Fail the test if the export instead contains outputs[0].name = analyse.cohortes_clients where an identifier or protocol value was expected. Run the field-specific assertion: compare resulting edge list from original versus target locale. This matters because business teams may confuse new cohorts with source customer data. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation output labels can be pleasant without promising the metric was validated into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify output dataset name, but it must retain the source evidence and must pass this field-specific test: compare resulting edge list from original versus target locale. Transfer destination: a business glossary card.

Schema, description, source systems and time

Module 11: Inputs versus outputs arrays

In a transformation joining invoices to payments, the source fixture is inputs = [raw.invoices, raw.payments]; outputs = [mart.settlements]. The tempting mistranslation is an illustration reversing the arrow from settlements to invoices, whereas the controlled explanation is “Invoices and payments feed the settlement mart”. This distinction matters because the direction of the dependency is part of the factual claim. A person who does not know the domain can easily overlook that an event boundary alone may not prove every input field feeds every output field. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 061: Catalog label versus identifier

A reader opens a catalog card for a transformation joining invoices to payments. The underlying evidence is inputs = [raw.invoices, raw.payments]; outputs = [mart.settlements]. The dangerous shortcut is an illustration reversing the arrow from settlements to invoices, because it treats a localized display convention as if it were the object itself. A better presentation is “Invoices and payments feed the settlement mart”. The primary contract is: the direction of the dependency is part of the factual claim. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, export graph edges and compare endpoint identities and direction. The language editor should also note that an event boundary alone may not prove every input field feeds every output field. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 062: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a transformation joining invoices to payments as ordinary text. It may output an illustration reversing the arrow from settlements to invoices, a grammatically plausible change that does not respect the protocol role of inputs = [raw.invoices, raw.payments]; outputs = [mart.settlements]. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Invoices and payments feed the settlement mart”. For this field the reason is not stylistic preference: the direction of the dependency is part of the factual claim. A reviewer should apply the concrete test to this example: export graph edges and compare endpoint identities and direction. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that an event boundary alone may not prove every input field feeds every output field.

Diagnostic 063: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a transformation joining invoices to payments; the resulting translation resembles an illustration reversing the arrow from settlements to invoices. Reconstruct the full path and the scenario, a transformation joining invoices to payments, before rewriting the label. Put the original evidence, inputs = [raw.invoices, raw.payments]; outputs = [mart.settlements], next to the human recommendation, “Invoices and payments feed the settlement mart”. Then ask whether the source record actually supports the claim: the direction of the dependency is part of the factual claim. A correct diagnosis is reproducible because export graph edges and compare endpoint identities and direction. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: an event boundary alone may not prove every input field feeds every output field. The goal is reader understanding without inventing a technical fact.

Diagnostic 064: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a transformation joining invoices to payments. If the translation has followed the shortcut an illustration reversing the arrow from settlements to invoices, the operational danger is that an incident investigation may work downstream instead of upstream. A better answer begins with the source evidence inputs = [raw.invoices, raw.payments]; outputs = [mart.settlements], describes it as “Invoices and payments feed the settlement mart”, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: export graph edges and compare endpoint identities and direction. The caveat an event boundary alone may not prove every input field feeds every output field also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 065: Bilingual reviewer handoff

Give an independent domain reviewer the source item inputs = [raw.invoices, raw.payments]; outputs = [mart.settlements], the rejected wording an illustration reversing the arrow from settlements to invoices, and the proposed user-facing wording “Invoices and payments feed the settlement mart”. Ask the reviewer to explain in plain language why the direction of the dependency is part of the factual claim. They should supply an example from a transformation joining invoices to payments, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: export graph edges and compare endpoint identities and direction. A final note should state an event boundary alone may not prove every input field feeds every output field, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 066: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a transformation joining invoices to payments. The technical reference inputs = [raw.invoices, raw.payments]; outputs = [mart.settlements] must still identify the same object or evidence, while the localized display may read “Invoices and payments feed the settlement mart”. Fail the test if the export instead contains an illustration reversing the arrow from settlements to invoices where an identifier or protocol value was expected. Run the field-specific assertion: export graph edges and compare endpoint identities and direction. This matters because an incident investigation may work downstream instead of upstream. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation an event boundary alone may not prove every input field feeds every output field into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify inputs versus outputs arrays, but it must retain the source evidence and must pass this field-specific test: export graph edges and compare endpoint identities and direction. Transfer destination: a dependency explorer with arrows.

Module 12: Dataset schema field key

In a nested customer profile table, the source fixture is schema.fields[].name = preferred_language. The tempting mistranslation is schema.fields[].name = langue_préférée, whereas the controlled explanation is “Preferred language” as a field label shown beside preferred_language. This distinction matters because field name is a structured schema identity, distinct from a description. A person who does not know the domain can easily overlook that nested field paths require their own identity and disambiguation. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 067: Catalog label versus identifier

A reader opens a catalog card for a nested customer profile table. The underlying evidence is schema.fields[].name = preferred_language. The dangerous shortcut is schema.fields[].name = langue_préférée, because it treats a localized display convention as if it were the object itself. A better presentation is “Preferred language” as a field label shown beside preferred_language. The primary contract is: field name is a structured schema identity, distinct from a description. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, query the raw schema facet and compare the exact field list. The language editor should also note that nested field paths require their own identity and disambiguation. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 068: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a nested customer profile table as ordinary text. It may output schema.fields[].name = langue_préférée, a grammatically plausible change that does not respect the protocol role of schema.fields[].name = preferred_language. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Preferred language” as a field label shown beside preferred_language. For this field the reason is not stylistic preference: field name is a structured schema identity, distinct from a description. A reviewer should apply the concrete test to this example: query the raw schema facet and compare the exact field list. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that nested field paths require their own identity and disambiguation.

Diagnostic 069: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a nested customer profile table; the resulting translation resembles schema.fields[].name = langue_préférée. Reconstruct the full path and the scenario, a nested customer profile table, before rewriting the label. Put the original evidence, schema.fields[].name = preferred_language, next to the human recommendation, “Preferred language” as a field label shown beside preferred_language. Then ask whether the source record actually supports the claim: field name is a structured schema identity, distinct from a description. A correct diagnosis is reproducible because query the raw schema facet and compare the exact field list. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: nested field paths require their own identity and disambiguation. The goal is reader understanding without inventing a technical fact.

Diagnostic 070: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a nested customer profile table. If the translation has followed the shortcut schema.fields[].name = langue_préférée, the operational danger is that an automated column match may fail after localization. A better answer begins with the source evidence schema.fields[].name = preferred_language, describes it as “Preferred language” as a field label shown beside preferred_language, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: query the raw schema facet and compare the exact field list. The caveat nested field paths require their own identity and disambiguation also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 071: Bilingual reviewer handoff

Give an independent domain reviewer the source item schema.fields[].name = preferred_language, the rejected wording schema.fields[].name = langue_préférée, and the proposed user-facing wording “Preferred language” as a field label shown beside preferred_language. Ask the reviewer to explain in plain language why field name is a structured schema identity, distinct from a description. They should supply an example from a nested customer profile table, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: query the raw schema facet and compare the exact field list. A final note should state nested field paths require their own identity and disambiguation, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 072: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a nested customer profile table. The technical reference schema.fields[].name = preferred_language must still identify the same object or evidence, while the localized display may read “Preferred language” as a field label shown beside preferred_language. Fail the test if the export instead contains schema.fields[].name = langue_préférée where an identifier or protocol value was expected. Run the field-specific assertion: query the raw schema facet and compare the exact field list. This matters because an automated column match may fail after localization. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation nested field paths require their own identity and disambiguation into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify dataset schema field key, but it must retain the source evidence and must pass this field-specific test: query the raw schema facet and compare the exact field list. Transfer destination: field lists and data dictionaries.

Module 13: Dataset schema type

In a field containing an integer account count, the source fixture is schema.fields[].type = integer. The tempting mistranslation is schema.fields[].type = number-with-text-label, whereas the controlled explanation is “Whole-number count” as descriptive help while integer remains the type. This distinction matters because types affect parsing and permissible operations; translation cannot change them. A person who does not know the domain can easily overlook that a field can be nullable without changing its declared base type. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 073: Catalog label versus identifier

A reader opens a catalog card for a field containing an integer account count. The underlying evidence is schema.fields[].type = integer. The dangerous shortcut is schema.fields[].type = number-with-text-label, because it treats a localized display convention as if it were the object itself. A better presentation is “Whole-number count” as descriptive help while integer remains the type. The primary contract is: types affect parsing and permissible operations; translation cannot change them. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, run a schema checker against the original and localized export. The language editor should also note that a field can be nullable without changing its declared base type. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 074: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a field containing an integer account count as ordinary text. It may output schema.fields[].type = number-with-text-label, a grammatically plausible change that does not respect the protocol role of schema.fields[].type = integer. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Whole-number count” as descriptive help while integer remains the type. For this field the reason is not stylistic preference: types affect parsing and permissible operations; translation cannot change them. A reviewer should apply the concrete test to this example: run a schema checker against the original and localized export. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that a field can be nullable without changing its declared base type.

Diagnostic 075: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a field containing an integer account count; the resulting translation resembles schema.fields[].type = number-with-text-label. Reconstruct the full path and the scenario, a field containing an integer account count, before rewriting the label. Put the original evidence, schema.fields[].type = integer, next to the human recommendation, “Whole-number count” as descriptive help while integer remains the type. Then ask whether the source record actually supports the claim: types affect parsing and permissible operations; translation cannot change them. A correct diagnosis is reproducible because run a schema checker against the original and localized export. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: a field can be nullable without changing its declared base type. The goal is reader understanding without inventing a technical fact.

Diagnostic 076: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a field containing an integer account count. If the translation has followed the shortcut schema.fields[].type = number-with-text-label, the operational danger is that a reader could expect decimals or free text where none are valid. A better answer begins with the source evidence schema.fields[].type = integer, describes it as “Whole-number count” as descriptive help while integer remains the type, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: run a schema checker against the original and localized export. The caveat a field can be nullable without changing its declared base type also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 077: Bilingual reviewer handoff

Give an independent domain reviewer the source item schema.fields[].type = integer, the rejected wording schema.fields[].type = number-with-text-label, and the proposed user-facing wording “Whole-number count” as descriptive help while integer remains the type. Ask the reviewer to explain in plain language why types affect parsing and permissible operations; translation cannot change them. They should supply an example from a field containing an integer account count, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: run a schema checker against the original and localized export. A final note should state a field can be nullable without changing its declared base type, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 078: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a field containing an integer account count. The technical reference schema.fields[].type = integer must still identify the same object or evidence, while the localized display may read “Whole-number count” as descriptive help while integer remains the type. Fail the test if the export instead contains schema.fields[].type = number-with-text-label where an identifier or protocol value was expected. Run the field-specific assertion: run a schema checker against the original and localized export. This matters because a reader could expect decimals or free text where none are valid. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation a field can be nullable without changing its declared base type into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify dataset schema type, but it must retain the source evidence and must pass this field-specific test: run a schema checker against the original and localized export. Transfer destination: analytics schema documentation.

Module 14: Field description prose

In a revenue field containing taxes and returns, the source fixture is schema.fields[].description = Net amount after returns, before tax. The tempting mistranslation is “Net amount after tax”, whereas the controlled explanation is “Net amount after returns, before tax” translated without changing qualifiers. This distinction matters because description is translatable prose, but its economic definition is constrained by source rules. A person who does not know the domain can easily overlook that a word such as net may hide multiple deductions in different organizations. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 079: Catalog label versus identifier

A reader opens a catalog card for a revenue field containing taxes and returns. The underlying evidence is schema.fields[].description = Net amount after returns, before tax. The dangerous shortcut is “Net amount after tax”, because it treats a localized display convention as if it were the object itself. A better presentation is “Net amount after returns, before tax” translated without changing qualifiers. The primary contract is: description is translatable prose, but its economic definition is constrained by source rules. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, ask a finance owner to approve the two qualifiers and inspect calculation logic. The language editor should also note that a word such as net may hide multiple deductions in different organizations. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 080: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a revenue field containing taxes and returns as ordinary text. It may output “Net amount after tax”, a grammatically plausible change that does not respect the protocol role of schema.fields[].description = Net amount after returns, before tax. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Net amount after returns, before tax” translated without changing qualifiers. For this field the reason is not stylistic preference: description is translatable prose, but its economic definition is constrained by source rules. A reviewer should apply the concrete test to this example: ask a finance owner to approve the two qualifiers and inspect calculation logic. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that a word such as net may hide multiple deductions in different organizations.

Diagnostic 081: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a revenue field containing taxes and returns; the resulting translation resembles “Net amount after tax”. Reconstruct the full path and the scenario, a revenue field containing taxes and returns, before rewriting the label. Put the original evidence, schema.fields[].description = Net amount after returns, before tax, next to the human recommendation, “Net amount after returns, before tax” translated without changing qualifiers. Then ask whether the source record actually supports the claim: description is translatable prose, but its economic definition is constrained by source rules. A correct diagnosis is reproducible because ask a finance owner to approve the two qualifiers and inspect calculation logic. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: a word such as net may hide multiple deductions in different organizations. The goal is reader understanding without inventing a technical fact.

Diagnostic 082: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a revenue field containing taxes and returns. If the translation has followed the shortcut “Net amount after tax”, the operational danger is that a translation could alter financial interpretation while all numbers stay fixed. A better answer begins with the source evidence schema.fields[].description = Net amount after returns, before tax, describes it as “Net amount after returns, before tax” translated without changing qualifiers, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: ask a finance owner to approve the two qualifiers and inspect calculation logic. The caveat a word such as net may hide multiple deductions in different organizations also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 083: Bilingual reviewer handoff

Give an independent domain reviewer the source item schema.fields[].description = Net amount after returns, before tax, the rejected wording “Net amount after tax”, and the proposed user-facing wording “Net amount after returns, before tax” translated without changing qualifiers. Ask the reviewer to explain in plain language why description is translatable prose, but its economic definition is constrained by source rules. They should supply an example from a revenue field containing taxes and returns, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: ask a finance owner to approve the two qualifiers and inspect calculation logic. A final note should state a word such as net may hide multiple deductions in different organizations, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 084: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a revenue field containing taxes and returns. The technical reference schema.fields[].description = Net amount after returns, before tax must still identify the same object or evidence, while the localized display may read “Net amount after returns, before tax” translated without changing qualifiers. Fail the test if the export instead contains “Net amount after tax” where an identifier or protocol value was expected. Run the field-specific assertion: ask a finance owner to approve the two qualifiers and inspect calculation logic. This matters because a translation could alter financial interpretation while all numbers stay fixed. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation a word such as net may hide multiple deductions in different organizations into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify field description prose, but it must retain the source evidence and must pass this field-specific test: ask a finance owner to approve the two qualifiers and inspect calculation logic. Transfer destination: metric glossary entries.

Module 15: Job documentation facet

In a pipeline producing aggregated usage reports, the source fixture is job.facets.documentation.description = Aggregates daily active sessions. The tempting mistranslation is “Counts all registered users”, whereas the controlled explanation is “Aggregates daily active sessions” with documented inclusion criteria. This distinction matters because a documentation string may be translated but does not supersede observed lineage. A person who does not know the domain can easily overlook that text/markdown content may contain code spans or links that must be preserved. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 085: Catalog label versus identifier

A reader opens a catalog card for a pipeline producing aggregated usage reports. The underlying evidence is job.facets.documentation.description = Aggregates daily active sessions. The dangerous shortcut is “Counts all registered users”, because it treats a localized display convention as if it were the object itself. A better presentation is “Aggregates daily active sessions” with documented inclusion criteria. The primary contract is: a documentation string may be translated but does not supersede observed lineage. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, check job source, owner description and known eligibility filters. The language editor should also note that text/markdown content may contain code spans or links that must be preserved. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 086: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a pipeline producing aggregated usage reports as ordinary text. It may output “Counts all registered users”, a grammatically plausible change that does not respect the protocol role of job.facets.documentation.description = Aggregates daily active sessions. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Aggregates daily active sessions” with documented inclusion criteria. For this field the reason is not stylistic preference: a documentation string may be translated but does not supersede observed lineage. A reviewer should apply the concrete test to this example: check job source, owner description and known eligibility filters. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that text/markdown content may contain code spans or links that must be preserved.

Diagnostic 087: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a pipeline producing aggregated usage reports; the resulting translation resembles “Counts all registered users”. Reconstruct the full path and the scenario, a pipeline producing aggregated usage reports, before rewriting the label. Put the original evidence, job.facets.documentation.description = Aggregates daily active sessions, next to the human recommendation, “Aggregates daily active sessions” with documented inclusion criteria. Then ask whether the source record actually supports the claim: a documentation string may be translated but does not supersede observed lineage. A correct diagnosis is reproducible because check job source, owner description and known eligibility filters. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: text/markdown content may contain code spans or links that must be preserved. The goal is reader understanding without inventing a technical fact.

Diagnostic 088: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a pipeline producing aggregated usage reports. If the translation has followed the shortcut “Counts all registered users”, the operational danger is that a translated README can misrepresent the actual pipeline. A better answer begins with the source evidence job.facets.documentation.description = Aggregates daily active sessions, describes it as “Aggregates daily active sessions” with documented inclusion criteria, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: check job source, owner description and known eligibility filters. The caveat text/markdown content may contain code spans or links that must be preserved also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 089: Bilingual reviewer handoff

Give an independent domain reviewer the source item job.facets.documentation.description = Aggregates daily active sessions, the rejected wording “Counts all registered users”, and the proposed user-facing wording “Aggregates daily active sessions” with documented inclusion criteria. Ask the reviewer to explain in plain language why a documentation string may be translated but does not supersede observed lineage. They should supply an example from a pipeline producing aggregated usage reports, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: check job source, owner description and known eligibility filters. A final note should state text/markdown content may contain code spans or links that must be preserved, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 090: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a pipeline producing aggregated usage reports. The technical reference job.facets.documentation.description = Aggregates daily active sessions must still identify the same object or evidence, while the localized display may read “Aggregates daily active sessions” with documented inclusion criteria. Fail the test if the export instead contains “Counts all registered users” where an identifier or protocol value was expected. Run the field-specific assertion: check job source, owner description and known eligibility filters. This matters because a translated README can misrepresent the actual pipeline. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation text/markdown content may contain code spans or links that must be preserved into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify job documentation facet, but it must retain the source evidence and must pass this field-specific test: check job source, owner description and known eligibility filters. Transfer destination: catalog documentation panels.

Module 16: SQL job facet

In a query filtering cancelled purchases, the source fixture is job.facets.sql.query = SELECT total FROM orders WHERE cancelled = false. The tempting mistranslation is SQL keywords and column names translated in-place, whereas the controlled explanation is “Reads the total for non-cancelled orders” outside the query. This distinction matters because SQL is executable or evidence-bearing syntax, not a translation source to rewrite. A person who does not know the domain can easily overlook that quoted literals are data; translating their content can change filtering semantics. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 091: Catalog label versus identifier

A reader opens a catalog card for a query filtering cancelled purchases. The underlying evidence is job.facets.sql.query = SELECT total FROM orders WHERE cancelled = false. The dangerous shortcut is SQL keywords and column names translated in-place, because it treats a localized display convention as if it were the object itself. A better presentation is “Reads the total for non-cancelled orders” outside the query. The primary contract is: SQL is executable or evidence-bearing syntax, not a translation source to rewrite. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, compare query hash and source text against the event’s SQL facet. The language editor should also note that quoted literals are data; translating their content can change filtering semantics. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 092: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a query filtering cancelled purchases as ordinary text. It may output SQL keywords and column names translated in-place, a grammatically plausible change that does not respect the protocol role of job.facets.sql.query = SELECT total FROM orders WHERE cancelled = false. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Reads the total for non-cancelled orders” outside the query. For this field the reason is not stylistic preference: SQL is executable or evidence-bearing syntax, not a translation source to rewrite. A reviewer should apply the concrete test to this example: compare query hash and source text against the event’s SQL facet. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that quoted literals are data; translating their content can change filtering semantics.

Diagnostic 093: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a query filtering cancelled purchases; the resulting translation resembles SQL keywords and column names translated in-place. Reconstruct the full path and the scenario, a query filtering cancelled purchases, before rewriting the label. Put the original evidence, job.facets.sql.query = SELECT total FROM orders WHERE cancelled = false, next to the human recommendation, “Reads the total for non-cancelled orders” outside the query. Then ask whether the source record actually supports the claim: SQL is executable or evidence-bearing syntax, not a translation source to rewrite. A correct diagnosis is reproducible because compare query hash and source text against the event’s SQL facet. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: quoted literals are data; translating their content can change filtering semantics. The goal is reader understanding without inventing a technical fact.

Diagnostic 094: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a query filtering cancelled purchases. If the translation has followed the shortcut SQL keywords and column names translated in-place, the operational danger is that a tiny word substitution may change the query or render it invalid. A better answer begins with the source evidence job.facets.sql.query = SELECT total FROM orders WHERE cancelled = false, describes it as “Reads the total for non-cancelled orders” outside the query, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: compare query hash and source text against the event’s SQL facet. The caveat quoted literals are data; translating their content can change filtering semantics also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 095: Bilingual reviewer handoff

Give an independent domain reviewer the source item job.facets.sql.query = SELECT total FROM orders WHERE cancelled = false, the rejected wording SQL keywords and column names translated in-place, and the proposed user-facing wording “Reads the total for non-cancelled orders” outside the query. Ask the reviewer to explain in plain language why SQL is executable or evidence-bearing syntax, not a translation source to rewrite. They should supply an example from a query filtering cancelled purchases, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: compare query hash and source text against the event’s SQL facet. A final note should state quoted literals are data; translating their content can change filtering semantics, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 096: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a query filtering cancelled purchases. The technical reference job.facets.sql.query = SELECT total FROM orders WHERE cancelled = false must still identify the same object or evidence, while the localized display may read “Reads the total for non-cancelled orders” outside the query. Fail the test if the export instead contains SQL keywords and column names translated in-place where an identifier or protocol value was expected. Run the field-specific assertion: compare query hash and source text against the event’s SQL facet. This matters because a tiny word substitution may change the query or render it invalid. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation quoted literals are data; translating their content can change filtering semantics into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify sql job facet, but it must retain the source evidence and must pass this field-specific test: compare query hash and source text against the event’s SQL facet. Transfer destination: developer-facing query explanations.

Module 17: Nominal time facet

In a schedule targeting each midnight window, the source fixture is run.facets.nominalTime.nominalStartTime = 2026-10-06T00:00:00Z. The tempting mistranslation is “Job executed precisely at midnight”, whereas the controlled explanation is “Scheduled window begins at 00:00 UTC”. This distinction matters because nominal time represents an expected scheduling window and can differ from eventTime. A person who does not know the domain can easily overlook that daily logical intervals differ from wall-clock execution moments. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 097: Catalog label versus identifier

A reader opens a catalog card for a schedule targeting each midnight window. The underlying evidence is run.facets.nominalTime.nominalStartTime = 2026-10-06T00:00:00Z. The dangerous shortcut is “Job executed precisely at midnight”, because it treats a localized display convention as if it were the object itself. A better presentation is “Scheduled window begins at 00:00 UTC”. The primary contract is: nominal time represents an expected scheduling window and can differ from eventTime. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, place nominalStartTime beside START eventTime on the same timeline. The language editor should also note that daily logical intervals differ from wall-clock execution moments. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 098: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a schedule targeting each midnight window as ordinary text. It may output “Job executed precisely at midnight”, a grammatically plausible change that does not respect the protocol role of run.facets.nominalTime.nominalStartTime = 2026-10-06T00:00:00Z. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Scheduled window begins at 00:00 UTC”. For this field the reason is not stylistic preference: nominal time represents an expected scheduling window and can differ from eventTime. A reviewer should apply the concrete test to this example: place nominalStartTime beside START eventTime on the same timeline. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that daily logical intervals differ from wall-clock execution moments.

Diagnostic 099: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a schedule targeting each midnight window; the resulting translation resembles “Job executed precisely at midnight”. Reconstruct the full path and the scenario, a schedule targeting each midnight window, before rewriting the label. Put the original evidence, run.facets.nominalTime.nominalStartTime = 2026-10-06T00:00:00Z, next to the human recommendation, “Scheduled window begins at 00:00 UTC”. Then ask whether the source record actually supports the claim: nominal time represents an expected scheduling window and can differ from eventTime. A correct diagnosis is reproducible because place nominalStartTime beside START eventTime on the same timeline. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: daily logical intervals differ from wall-clock execution moments. The goal is reader understanding without inventing a technical fact.

Diagnostic 100: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a schedule targeting each midnight window. If the translation has followed the shortcut “Job executed precisely at midnight”, the operational danger is that late execution may be hidden by a fluent but untrue translation. A better answer begins with the source evidence run.facets.nominalTime.nominalStartTime = 2026-10-06T00:00:00Z, describes it as “Scheduled window begins at 00:00 UTC”, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: place nominalStartTime beside START eventTime on the same timeline. The caveat daily logical intervals differ from wall-clock execution moments also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 101: Bilingual reviewer handoff

Give an independent domain reviewer the source item run.facets.nominalTime.nominalStartTime = 2026-10-06T00:00:00Z, the rejected wording “Job executed precisely at midnight”, and the proposed user-facing wording “Scheduled window begins at 00:00 UTC”. Ask the reviewer to explain in plain language why nominal time represents an expected scheduling window and can differ from eventTime. They should supply an example from a schedule targeting each midnight window, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: place nominalStartTime beside START eventTime on the same timeline. A final note should state daily logical intervals differ from wall-clock execution moments, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 102: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a schedule targeting each midnight window. The technical reference run.facets.nominalTime.nominalStartTime = 2026-10-06T00:00:00Z must still identify the same object or evidence, while the localized display may read “Scheduled window begins at 00:00 UTC”. Fail the test if the export instead contains “Job executed precisely at midnight” where an identifier or protocol value was expected. Run the field-specific assertion: place nominalStartTime beside START eventTime on the same timeline. This matters because late execution may be hidden by a fluent but untrue translation. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation daily logical intervals differ from wall-clock execution moments into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify nominal time facet, but it must retain the source evidence and must pass this field-specific test: place nominalStartTime beside START eventTime on the same timeline. Transfer destination: an SLA explanation.

Module 18: Parent run facet

In a workflow that spawns three child tasks, the source fixture is run.facets.parent.run.runId = parent-run-uuid. The tempting mistranslation is “This job owns the data”, whereas the controlled explanation is “Child task of the parent execution” with unchanged parent run reference. This distinction matters because parent links express execution hierarchy, not ownership or data derivation. A person who does not know the domain can easily overlook that parentage does not necessarily prove direct dataset lineage. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 103: Catalog label versus identifier

A reader opens a catalog card for a workflow that spawns three child tasks. The underlying evidence is run.facets.parent.run.runId = parent-run-uuid. The dangerous shortcut is “This job owns the data”, because it treats a localized display convention as if it were the object itself. A better presentation is “Child task of the parent execution” with unchanged parent run reference. The primary contract is: parent links express execution hierarchy, not ownership or data derivation. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, trace the parent run ID to the orchestrator event chain. The language editor should also note that parentage does not necessarily prove direct dataset lineage. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 104: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a workflow that spawns three child tasks as ordinary text. It may output “This job owns the data”, a grammatically plausible change that does not respect the protocol role of run.facets.parent.run.runId = parent-run-uuid. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Child task of the parent execution” with unchanged parent run reference. For this field the reason is not stylistic preference: parent links express execution hierarchy, not ownership or data derivation. A reviewer should apply the concrete test to this example: trace the parent run ID to the orchestrator event chain. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that parentage does not necessarily prove direct dataset lineage.

Diagnostic 105: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a workflow that spawns three child tasks; the resulting translation resembles “This job owns the data”. Reconstruct the full path and the scenario, a workflow that spawns three child tasks, before rewriting the label. Put the original evidence, run.facets.parent.run.runId = parent-run-uuid, next to the human recommendation, “Child task of the parent execution” with unchanged parent run reference. Then ask whether the source record actually supports the claim: parent links express execution hierarchy, not ownership or data derivation. A correct diagnosis is reproducible because trace the parent run ID to the orchestrator event chain. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: parentage does not necessarily prove direct dataset lineage. The goal is reader understanding without inventing a technical fact.

Diagnostic 106: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a workflow that spawns three child tasks. If the translation has followed the shortcut “This job owns the data”, the operational danger is that a nested task might be mistaken for the root pipeline. A better answer begins with the source evidence run.facets.parent.run.runId = parent-run-uuid, describes it as “Child task of the parent execution” with unchanged parent run reference, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: trace the parent run ID to the orchestrator event chain. The caveat parentage does not necessarily prove direct dataset lineage also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 107: Bilingual reviewer handoff

Give an independent domain reviewer the source item run.facets.parent.run.runId = parent-run-uuid, the rejected wording “This job owns the data”, and the proposed user-facing wording “Child task of the parent execution” with unchanged parent run reference. Ask the reviewer to explain in plain language why parent links express execution hierarchy, not ownership or data derivation. They should supply an example from a workflow that spawns three child tasks, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: trace the parent run ID to the orchestrator event chain. A final note should state parentage does not necessarily prove direct dataset lineage, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 108: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a workflow that spawns three child tasks. The technical reference run.facets.parent.run.runId = parent-run-uuid must still identify the same object or evidence, while the localized display may read “Child task of the parent execution” with unchanged parent run reference. Fail the test if the export instead contains “This job owns the data” where an identifier or protocol value was expected. Run the field-specific assertion: trace the parent run ID to the orchestrator event chain. This matters because a nested task might be mistaken for the root pipeline. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation parentage does not necessarily prove direct dataset lineage into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify parent run facet, but it must retain the source evidence and must pass this field-specific test: trace the parent run ID to the orchestrator event chain. Transfer destination: a workflow hierarchy viewer.

Module 19: Processing engine facet

In a transformation executed by a distributed engine, the source fixture is run.facets.processing_engine-like metadata = Spark 4.x. The tempting mistranslation is “Spark” interpreted as a person or a dataset, whereas the controlled explanation is “Processed with the Spark engine” as contextual description. This distinction matters because processing engine describes execution technology, not a lineage edge. A person who does not know the domain can easily overlook that do not assert a nonstandard key path without checking the emitter’s facet schema. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 109: Catalog label versus identifier

A reader opens a catalog card for a transformation executed by a distributed engine. The underlying evidence is run.facets.processing_engine-like metadata = Spark 4.x. The dangerous shortcut is “Spark” interpreted as a person or a dataset, because it treats a localized display convention as if it were the object itself. A better presentation is “Processed with the Spark engine” as contextual description. The primary contract is: processing engine describes execution technology, not a lineage edge. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, inspect actual facet schema and producer documentation for the emitter. The language editor should also note that do not assert a nonstandard key path without checking the emitter’s facet schema. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 110: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a transformation executed by a distributed engine as ordinary text. It may output “Spark” interpreted as a person or a dataset, a grammatically plausible change that does not respect the protocol role of run.facets.processing_engine-like metadata = Spark 4.x. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Processed with the Spark engine” as contextual description. For this field the reason is not stylistic preference: processing engine describes execution technology, not a lineage edge. A reviewer should apply the concrete test to this example: inspect actual facet schema and producer documentation for the emitter. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that do not assert a nonstandard key path without checking the emitter’s facet schema.

Diagnostic 111: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a transformation executed by a distributed engine; the resulting translation resembles “Spark” interpreted as a person or a dataset. Reconstruct the full path and the scenario, a transformation executed by a distributed engine, before rewriting the label. Put the original evidence, run.facets.processing_engine-like metadata = Spark 4.x, next to the human recommendation, “Processed with the Spark engine” as contextual description. Then ask whether the source record actually supports the claim: processing engine describes execution technology, not a lineage edge. A correct diagnosis is reproducible because inspect actual facet schema and producer documentation for the emitter. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: do not assert a nonstandard key path without checking the emitter’s facet schema. The goal is reader understanding without inventing a technical fact.

Diagnostic 112: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a transformation executed by a distributed engine. If the translation has followed the shortcut “Spark” interpreted as a person or a dataset, the operational danger is that troubleshooting may be routed to the wrong runtime. A better answer begins with the source evidence run.facets.processing_engine-like metadata = Spark 4.x, describes it as “Processed with the Spark engine” as contextual description, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: inspect actual facet schema and producer documentation for the emitter. The caveat do not assert a nonstandard key path without checking the emitter’s facet schema also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 113: Bilingual reviewer handoff

Give an independent domain reviewer the source item run.facets.processing_engine-like metadata = Spark 4.x, the rejected wording “Spark” interpreted as a person or a dataset, and the proposed user-facing wording “Processed with the Spark engine” as contextual description. Ask the reviewer to explain in plain language why processing engine describes execution technology, not a lineage edge. They should supply an example from a transformation executed by a distributed engine, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: inspect actual facet schema and producer documentation for the emitter. A final note should state do not assert a nonstandard key path without checking the emitter’s facet schema, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 114: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a transformation executed by a distributed engine. The technical reference run.facets.processing_engine-like metadata = Spark 4.x must still identify the same object or evidence, while the localized display may read “Processed with the Spark engine” as contextual description. Fail the test if the export instead contains “Spark” interpreted as a person or a dataset where an identifier or protocol value was expected. Run the field-specific assertion: inspect actual facet schema and producer documentation for the emitter. This matters because troubleshooting may be routed to the wrong runtime. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation do not assert a nonstandard key path without checking the emitter’s facet schema into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify processing engine facet, but it must retain the source evidence and must pass this field-specific test: inspect actual facet schema and producer documentation for the emitter. Transfer destination: an incident technical-details panel.

Module 20: DataSource dataset facet

In a table inside a managed data warehouse, the source fixture is dataset.facets.dataSource.name = warehouse-01. The tempting mistranslation is data source name translated into an imaginary new database, whereas the controlled explanation is “Main analytics warehouse” as a catalog alias. This distinction matters because a data source facet identifies the containing system or location. A person who does not know the domain can easily overlook that display descriptions should not leak restricted endpoints. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 115: Catalog label versus identifier

A reader opens a catalog card for a table inside a managed data warehouse. The underlying evidence is dataset.facets.dataSource.name = warehouse-01. The dangerous shortcut is data source name translated into an imaginary new database, because it treats a localized display convention as if it were the object itself. A better presentation is “Main analytics warehouse” as a catalog alias. The primary contract is: a data source facet identifies the containing system or location. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, cross-check connection identity and dataset namespace. The language editor should also note that display descriptions should not leak restricted endpoints. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 116: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a table inside a managed data warehouse as ordinary text. It may output data source name translated into an imaginary new database, a grammatically plausible change that does not respect the protocol role of dataset.facets.dataSource.name = warehouse-01. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Main analytics warehouse” as a catalog alias. For this field the reason is not stylistic preference: a data source facet identifies the containing system or location. A reviewer should apply the concrete test to this example: cross-check connection identity and dataset namespace. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that display descriptions should not leak restricted endpoints.

Diagnostic 117: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a table inside a managed data warehouse; the resulting translation resembles data source name translated into an imaginary new database. Reconstruct the full path and the scenario, a table inside a managed data warehouse, before rewriting the label. Put the original evidence, dataset.facets.dataSource.name = warehouse-01, next to the human recommendation, “Main analytics warehouse” as a catalog alias. Then ask whether the source record actually supports the claim: a data source facet identifies the containing system or location. A correct diagnosis is reproducible because cross-check connection identity and dataset namespace. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: display descriptions should not leak restricted endpoints. The goal is reader understanding without inventing a technical fact.

Diagnostic 118: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a table inside a managed data warehouse. If the translation has followed the shortcut data source name translated into an imaginary new database, the operational danger is that audit readers may infer the wrong hosting system. A better answer begins with the source evidence dataset.facets.dataSource.name = warehouse-01, describes it as “Main analytics warehouse” as a catalog alias, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: cross-check connection identity and dataset namespace. The caveat display descriptions should not leak restricted endpoints also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 119: Bilingual reviewer handoff

Give an independent domain reviewer the source item dataset.facets.dataSource.name = warehouse-01, the rejected wording data source name translated into an imaginary new database, and the proposed user-facing wording “Main analytics warehouse” as a catalog alias. Ask the reviewer to explain in plain language why a data source facet identifies the containing system or location. They should supply an example from a table inside a managed data warehouse, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: cross-check connection identity and dataset namespace. A final note should state display descriptions should not leak restricted endpoints, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 120: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a table inside a managed data warehouse. The technical reference dataset.facets.dataSource.name = warehouse-01 must still identify the same object or evidence, while the localized display may read “Main analytics warehouse” as a catalog alias. Fail the test if the export instead contains data source name translated into an imaginary new database where an identifier or protocol value was expected. Run the field-specific assertion: cross-check connection identity and dataset namespace. This matters because audit readers may infer the wrong hosting system. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation display descriptions should not leak restricted endpoints into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify datasource dataset facet, but it must retain the source evidence and must pass this field-specific test: cross-check connection identity and dataset namespace. Transfer destination: a multi-warehouse governance catalog.

Module 21: Dataset version facet

In a snapshot-based analytical table, the source fixture is dataset.facets.version.datasetVersion = 90214. The tempting mistranslation is “Latest” used for this historical snapshot, whereas the controlled explanation is “Snapshot version 90214” and an explicit capture date. This distinction matters because a version token identifies a state, not a guarantee of recency. A person who does not know the domain can easily overlook that some stores use commit IDs rather than ordered version numbers. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 121: Catalog label versus identifier

A reader opens a catalog card for a snapshot-based analytical table. The underlying evidence is dataset.facets.version.datasetVersion = 90214. The dangerous shortcut is “Latest” used for this historical snapshot, because it treats a localized display convention as if it were the object itself. A better presentation is “Snapshot version 90214” and an explicit capture date. The primary contract is: a version token identifies a state, not a guarantee of recency. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, compare the snapshot reference used by the run and by the catalog. The language editor should also note that some stores use commit IDs rather than ordered version numbers. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 122: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a snapshot-based analytical table as ordinary text. It may output “Latest” used for this historical snapshot, a grammatically plausible change that does not respect the protocol role of dataset.facets.version.datasetVersion = 90214. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Snapshot version 90214” and an explicit capture date. For this field the reason is not stylistic preference: a version token identifies a state, not a guarantee of recency. A reviewer should apply the concrete test to this example: compare the snapshot reference used by the run and by the catalog. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that some stores use commit IDs rather than ordered version numbers.

Diagnostic 123: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a snapshot-based analytical table; the resulting translation resembles “Latest” used for this historical snapshot. Reconstruct the full path and the scenario, a snapshot-based analytical table, before rewriting the label. Put the original evidence, dataset.facets.version.datasetVersion = 90214, next to the human recommendation, “Snapshot version 90214” and an explicit capture date. Then ask whether the source record actually supports the claim: a version token identifies a state, not a guarantee of recency. A correct diagnosis is reproducible because compare the snapshot reference used by the run and by the catalog. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: some stores use commit IDs rather than ordered version numbers. The goal is reader understanding without inventing a technical fact.

Diagnostic 124: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a snapshot-based analytical table. If the translation has followed the shortcut “Latest” used for this historical snapshot, the operational danger is that a reviewer could compare metrics from different versions. A better answer begins with the source evidence dataset.facets.version.datasetVersion = 90214, describes it as “Snapshot version 90214” and an explicit capture date, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: compare the snapshot reference used by the run and by the catalog. The caveat some stores use commit IDs rather than ordered version numbers also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 125: Bilingual reviewer handoff

Give an independent domain reviewer the source item dataset.facets.version.datasetVersion = 90214, the rejected wording “Latest” used for this historical snapshot, and the proposed user-facing wording “Snapshot version 90214” and an explicit capture date. Ask the reviewer to explain in plain language why a version token identifies a state, not a guarantee of recency. They should supply an example from a snapshot-based analytical table, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: compare the snapshot reference used by the run and by the catalog. A final note should state some stores use commit IDs rather than ordered version numbers, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 126: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a snapshot-based analytical table. The technical reference dataset.facets.version.datasetVersion = 90214 must still identify the same object or evidence, while the localized display may read “Snapshot version 90214” and an explicit capture date. Fail the test if the export instead contains “Latest” used for this historical snapshot where an identifier or protocol value was expected. Run the field-specific assertion: compare the snapshot reference used by the run and by the catalog. This matters because a reviewer could compare metrics from different versions. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation some stores use commit IDs rather than ordered version numbers into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify dataset version facet, but it must retain the source evidence and must pass this field-specific test: compare the snapshot reference used by the run and by the catalog. Transfer destination: a reproducibility report.

Module 22: Lifecycle state change facet

In a source table being renamed during migration, the source fixture is dataset.facets.lifecycleStateChange.lifecycleStateChange = RENAME. The tempting mistranslation is “The dataset was deleted”, whereas the controlled explanation is “Renamed during the authorized migration” with old and new references. This distinction matters because lifecycle operations are controlled semantics, not interchangeable metaphors. A person who does not know the domain can easily overlook that rename, drop, overwrite and truncate require different consequences. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 127: Catalog label versus identifier

A reader opens a catalog card for a source table being renamed during migration. The underlying evidence is dataset.facets.lifecycleStateChange.lifecycleStateChange = RENAME. The dangerous shortcut is “The dataset was deleted”, because it treats a localized display convention as if it were the object itself. A better presentation is “Renamed during the authorized migration” with old and new references. The primary contract is: lifecycle operations are controlled semantics, not interchangeable metaphors. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, check change event and any affected symlinks or identifier mapping. The language editor should also note that rename, drop, overwrite and truncate require different consequences. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 128: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a source table being renamed during migration as ordinary text. It may output “The dataset was deleted”, a grammatically plausible change that does not respect the protocol role of dataset.facets.lifecycleStateChange.lifecycleStateChange = RENAME. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Renamed during the authorized migration” with old and new references. For this field the reason is not stylistic preference: lifecycle operations are controlled semantics, not interchangeable metaphors. A reviewer should apply the concrete test to this example: check change event and any affected symlinks or identifier mapping. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that rename, drop, overwrite and truncate require different consequences.

Diagnostic 129: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a source table being renamed during migration; the resulting translation resembles “The dataset was deleted”. Reconstruct the full path and the scenario, a source table being renamed during migration, before rewriting the label. Put the original evidence, dataset.facets.lifecycleStateChange.lifecycleStateChange = RENAME, next to the human recommendation, “Renamed during the authorized migration” with old and new references. Then ask whether the source record actually supports the claim: lifecycle operations are controlled semantics, not interchangeable metaphors. A correct diagnosis is reproducible because check change event and any affected symlinks or identifier mapping. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: rename, drop, overwrite and truncate require different consequences. The goal is reader understanding without inventing a technical fact.

Diagnostic 130: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a source table being renamed during migration. If the translation has followed the shortcut “The dataset was deleted”, the operational danger is that a false deletion notice may trigger unnecessary recovery. A better answer begins with the source evidence dataset.facets.lifecycleStateChange.lifecycleStateChange = RENAME, describes it as “Renamed during the authorized migration” with old and new references, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: check change event and any affected symlinks or identifier mapping. The caveat rename, drop, overwrite and truncate require different consequences also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 131: Bilingual reviewer handoff

Give an independent domain reviewer the source item dataset.facets.lifecycleStateChange.lifecycleStateChange = RENAME, the rejected wording “The dataset was deleted”, and the proposed user-facing wording “Renamed during the authorized migration” with old and new references. Ask the reviewer to explain in plain language why lifecycle operations are controlled semantics, not interchangeable metaphors. They should supply an example from a source table being renamed during migration, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: check change event and any affected symlinks or identifier mapping. A final note should state rename, drop, overwrite and truncate require different consequences, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 132: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a source table being renamed during migration. The technical reference dataset.facets.lifecycleStateChange.lifecycleStateChange = RENAME must still identify the same object or evidence, while the localized display may read “Renamed during the authorized migration” with old and new references. Fail the test if the export instead contains “The dataset was deleted” where an identifier or protocol value was expected. Run the field-specific assertion: check change event and any affected symlinks or identifier mapping. This matters because a false deletion notice may trigger unnecessary recovery. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation rename, drop, overwrite and truncate require different consequences into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify lifecycle state change facet, but it must retain the source evidence and must pass this field-specific test: check change event and any affected symlinks or identifier mapping. Transfer destination: a migration audit timeline.

Quality, transformations and graph causality

Module 23: Output statistics facet

In a transformed dataset reporting written records, the source fixture is outputs[0].outputFacets.outputStatistics.rowCount = 12045. The tempting mistranslation is “12,045 distinct customers”, whereas the controlled explanation is “12,045 output rows written” unless deduplication is independently proved. This distinction matters because output statistics describe a specific emitted result, not every business interpretation. A person who does not know the domain can easily overlook that rowCount and bytes convey different properties of an output. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 133: Catalog label versus identifier

A reader opens a catalog card for a transformed dataset reporting written records. The underlying evidence is outputs[0].outputFacets.outputStatistics.rowCount = 12045. The dangerous shortcut is “12,045 distinct customers”, because it treats a localized display convention as if it were the object itself. A better presentation is “12,045 output rows written” unless deduplication is independently proved. The primary contract is: output statistics describe a specific emitted result, not every business interpretation. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, verify measurement scope and time against output event. The language editor should also note that rowCount and bytes convey different properties of an output. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 134: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a transformed dataset reporting written records as ordinary text. It may output “12,045 distinct customers”, a grammatically plausible change that does not respect the protocol role of outputs[0].outputFacets.outputStatistics.rowCount = 12045. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “12,045 output rows written” unless deduplication is independently proved. For this field the reason is not stylistic preference: output statistics describe a specific emitted result, not every business interpretation. A reviewer should apply the concrete test to this example: verify measurement scope and time against output event. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that rowCount and bytes convey different properties of an output.

Diagnostic 135: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a transformed dataset reporting written records; the resulting translation resembles “12,045 distinct customers”. Reconstruct the full path and the scenario, a transformed dataset reporting written records, before rewriting the label. Put the original evidence, outputs[0].outputFacets.outputStatistics.rowCount = 12045, next to the human recommendation, “12,045 output rows written” unless deduplication is independently proved. Then ask whether the source record actually supports the claim: output statistics describe a specific emitted result, not every business interpretation. A correct diagnosis is reproducible because verify measurement scope and time against output event. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: rowCount and bytes convey different properties of an output. The goal is reader understanding without inventing a technical fact.

Diagnostic 136: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a transformed dataset reporting written records. If the translation has followed the shortcut “12,045 distinct customers”, the operational danger is that rows may be misrepresented as people or transactions. A better answer begins with the source evidence outputs[0].outputFacets.outputStatistics.rowCount = 12045, describes it as “12,045 output rows written” unless deduplication is independently proved, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: verify measurement scope and time against output event. The caveat rowCount and bytes convey different properties of an output also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 137: Bilingual reviewer handoff

Give an independent domain reviewer the source item outputs[0].outputFacets.outputStatistics.rowCount = 12045, the rejected wording “12,045 distinct customers”, and the proposed user-facing wording “12,045 output rows written” unless deduplication is independently proved. Ask the reviewer to explain in plain language why output statistics describe a specific emitted result, not every business interpretation. They should supply an example from a transformed dataset reporting written records, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: verify measurement scope and time against output event. A final note should state rowCount and bytes convey different properties of an output, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 138: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a transformed dataset reporting written records. The technical reference outputs[0].outputFacets.outputStatistics.rowCount = 12045 must still identify the same object or evidence, while the localized display may read “12,045 output rows written” unless deduplication is independently proved. Fail the test if the export instead contains “12,045 distinct customers” where an identifier or protocol value was expected. Run the field-specific assertion: verify measurement scope and time against output event. This matters because rows may be misrepresented as people or transactions. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation rowCount and bytes convey different properties of an output into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify output statistics facet, but it must retain the source evidence and must pass this field-specific test: verify measurement scope and time against output event. Transfer destination: an executive freshness card.

Module 24: Data quality metrics facet

In an input dataset with missing postal codes, the source fixture is dataset.facets.dataQualityMetrics.columnMetrics.postcode.nullCount = 39. The tempting mistranslation is “39 invalid addresses”, whereas the controlled explanation is “39 null postcode values in the measured dataset”. This distinction matters because nullCount is a measurable count; invalidity is a separate rule. A person who does not know the domain can easily overlook that null may mean unavailable, withheld or not applicable depending on schema. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 139: Catalog label versus identifier

A reader opens a catalog card for an input dataset with missing postal codes. The underlying evidence is dataset.facets.dataQualityMetrics.columnMetrics.postcode.nullCount = 39. The dangerous shortcut is “39 invalid addresses”, because it treats a localized display convention as if it were the object itself. A better presentation is “39 null postcode values in the measured dataset”. The primary contract is: nullCount is a measurable count; invalidity is a separate rule. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, inspect captureDate, metric scope, field identity and source rules. The language editor should also note that null may mean unavailable, withheld or not applicable depending on schema. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 140: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for an input dataset with missing postal codes as ordinary text. It may output “39 invalid addresses”, a grammatically plausible change that does not respect the protocol role of dataset.facets.dataQualityMetrics.columnMetrics.postcode.nullCount = 39. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “39 null postcode values in the measured dataset”. For this field the reason is not stylistic preference: nullCount is a measurable count; invalidity is a separate rule. A reviewer should apply the concrete test to this example: inspect captureDate, metric scope, field identity and source rules. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that null may mean unavailable, withheld or not applicable depending on schema.

Diagnostic 141: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of an input dataset with missing postal codes; the resulting translation resembles “39 invalid addresses”. Reconstruct the full path and the scenario, an input dataset with missing postal codes, before rewriting the label. Put the original evidence, dataset.facets.dataQualityMetrics.columnMetrics.postcode.nullCount = 39, next to the human recommendation, “39 null postcode values in the measured dataset”. Then ask whether the source record actually supports the claim: nullCount is a measurable count; invalidity is a separate rule. A correct diagnosis is reproducible because inspect captureDate, metric scope, field identity and source rules. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: null may mean unavailable, withheld or not applicable depending on schema. The goal is reader understanding without inventing a technical fact.

Diagnostic 142: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in an input dataset with missing postal codes. If the translation has followed the shortcut “39 invalid addresses”, the operational danger is that a defect report may exaggerate severity or accuse the wrong process. A better answer begins with the source evidence dataset.facets.dataQualityMetrics.columnMetrics.postcode.nullCount = 39, describes it as “39 null postcode values in the measured dataset”, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: inspect captureDate, metric scope, field identity and source rules. The caveat null may mean unavailable, withheld or not applicable depending on schema also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 143: Bilingual reviewer handoff

Give an independent domain reviewer the source item dataset.facets.dataQualityMetrics.columnMetrics.postcode.nullCount = 39, the rejected wording “39 invalid addresses”, and the proposed user-facing wording “39 null postcode values in the measured dataset”. Ask the reviewer to explain in plain language why nullCount is a measurable count; invalidity is a separate rule. They should supply an example from an input dataset with missing postal codes, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: inspect captureDate, metric scope, field identity and source rules. A final note should state null may mean unavailable, withheld or not applicable depending on schema, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 144: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for an input dataset with missing postal codes. The technical reference dataset.facets.dataQualityMetrics.columnMetrics.postcode.nullCount = 39 must still identify the same object or evidence, while the localized display may read “39 null postcode values in the measured dataset”. Fail the test if the export instead contains “39 invalid addresses” where an identifier or protocol value was expected. Run the field-specific assertion: inspect captureDate, metric scope, field identity and source rules. This matters because a defect report may exaggerate severity or accuse the wrong process. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation null may mean unavailable, withheld or not applicable depending on schema into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify data quality metrics facet, but it must retain the source evidence and must pass this field-specific test: inspect captureDate, metric scope, field identity and source rules. Transfer destination: a multilingual data-health report.

Module 25: Data quality assertions

In a rule requiring nonnegative stock counts, the source fixture is dataset facets report an assertion result for stock_count >= 0. The tempting mistranslation is “All inventory figures are accurate”, whereas the controlled explanation is “The nonnegative-stock assertion passed for the tested scope”. This distinction matters because passing one assertion is not proof of complete data quality. A person who does not know the domain can easily overlook that a failed assertion is not necessarily the same as a failed pipeline run. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 145: Catalog label versus identifier

A reader opens a catalog card for a rule requiring nonnegative stock counts. The underlying evidence is dataset facets report an assertion result for stock_count >= 0. The dangerous shortcut is “All inventory figures are accurate”, because it treats a localized display convention as if it were the object itself. A better presentation is “The nonnegative-stock assertion passed for the tested scope”. The primary contract is: passing one assertion is not proof of complete data quality. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, read the assertion definition, outcome and evaluated dataset version. The language editor should also note that a failed assertion is not necessarily the same as a failed pipeline run. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 146: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a rule requiring nonnegative stock counts as ordinary text. It may output “All inventory figures are accurate”, a grammatically plausible change that does not respect the protocol role of dataset facets report an assertion result for stock_count >= 0. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “The nonnegative-stock assertion passed for the tested scope”. For this field the reason is not stylistic preference: passing one assertion is not proof of complete data quality. A reviewer should apply the concrete test to this example: read the assertion definition, outcome and evaluated dataset version. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that a failed assertion is not necessarily the same as a failed pipeline run.

Diagnostic 147: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a rule requiring nonnegative stock counts; the resulting translation resembles “All inventory figures are accurate”. Reconstruct the full path and the scenario, a rule requiring nonnegative stock counts, before rewriting the label. Put the original evidence, dataset facets report an assertion result for stock_count >= 0, next to the human recommendation, “The nonnegative-stock assertion passed for the tested scope”. Then ask whether the source record actually supports the claim: passing one assertion is not proof of complete data quality. A correct diagnosis is reproducible because read the assertion definition, outcome and evaluated dataset version. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: a failed assertion is not necessarily the same as a failed pipeline run. The goal is reader understanding without inventing a technical fact.

Diagnostic 148: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a rule requiring nonnegative stock counts. If the translation has followed the shortcut “All inventory figures are accurate”, the operational danger is that a narrow test may be promoted into a misleading quality guarantee. A better answer begins with the source evidence dataset facets report an assertion result for stock_count >= 0, describes it as “The nonnegative-stock assertion passed for the tested scope”, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: read the assertion definition, outcome and evaluated dataset version. The caveat a failed assertion is not necessarily the same as a failed pipeline run also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 149: Bilingual reviewer handoff

Give an independent domain reviewer the source item dataset facets report an assertion result for stock_count >= 0, the rejected wording “All inventory figures are accurate”, and the proposed user-facing wording “The nonnegative-stock assertion passed for the tested scope”. Ask the reviewer to explain in plain language why passing one assertion is not proof of complete data quality. They should supply an example from a rule requiring nonnegative stock counts, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: read the assertion definition, outcome and evaluated dataset version. A final note should state a failed assertion is not necessarily the same as a failed pipeline run, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 150: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a rule requiring nonnegative stock counts. The technical reference dataset facets report an assertion result for stock_count >= 0 must still identify the same object or evidence, while the localized display may read “The nonnegative-stock assertion passed for the tested scope”. Fail the test if the export instead contains “All inventory figures are accurate” where an identifier or protocol value was expected. Run the field-specific assertion: read the assertion definition, outcome and evaluated dataset version. This matters because a narrow test may be promoted into a misleading quality guarantee. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation a failed assertion is not necessarily the same as a failed pipeline run into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify data quality assertions, but it must retain the source evidence and must pass this field-specific test: read the assertion definition, outcome and evaluated dataset version. Transfer destination: a stewardship certification page.

Module 26: Column lineage output field

In an output amount calculated from quantity and price, the source fixture is output column total_value uses input fields unit_price and quantity. The tempting mistranslation is “Total value is copied unchanged from price”, whereas the controlled explanation is “Total value is calculated from unit price and quantity”. This distinction matters because field lineage represents dependencies of an output field on its sources. A person who does not know the domain can easily overlook that a many-to-one calculation is not an identity mapping. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 151: Catalog label versus identifier

A reader opens a catalog card for an output amount calculated from quantity and price. The underlying evidence is output column total_value uses input fields unit_price and quantity. The dangerous shortcut is “Total value is copied unchanged from price”, because it treats a localized display convention as if it were the object itself. A better presentation is “Total value is calculated from unit price and quantity”. The primary contract is: field lineage represents dependencies of an output field on its sources. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, inspect inputFields for each output column and the query expression. The language editor should also note that a many-to-one calculation is not an identity mapping. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 152: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for an output amount calculated from quantity and price as ordinary text. It may output “Total value is copied unchanged from price”, a grammatically plausible change that does not respect the protocol role of output column total_value uses input fields unit_price and quantity. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Total value is calculated from unit price and quantity”. For this field the reason is not stylistic preference: field lineage represents dependencies of an output field on its sources. A reviewer should apply the concrete test to this example: inspect inputFields for each output column and the query expression. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that a many-to-one calculation is not an identity mapping.

Diagnostic 153: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of an output amount calculated from quantity and price; the resulting translation resembles “Total value is copied unchanged from price”. Reconstruct the full path and the scenario, an output amount calculated from quantity and price, before rewriting the label. Put the original evidence, output column total_value uses input fields unit_price and quantity, next to the human recommendation, “Total value is calculated from unit price and quantity”. Then ask whether the source record actually supports the claim: field lineage represents dependencies of an output field on its sources. A correct diagnosis is reproducible because inspect inputFields for each output column and the query expression. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: a many-to-one calculation is not an identity mapping. The goal is reader understanding without inventing a technical fact.

Diagnostic 154: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in an output amount calculated from quantity and price. If the translation has followed the shortcut “Total value is copied unchanged from price”, the operational danger is that a customer may misunderstand why totals changed. A better answer begins with the source evidence output column total_value uses input fields unit_price and quantity, describes it as “Total value is calculated from unit price and quantity”, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: inspect inputFields for each output column and the query expression. The caveat a many-to-one calculation is not an identity mapping also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 155: Bilingual reviewer handoff

Give an independent domain reviewer the source item output column total_value uses input fields unit_price and quantity, the rejected wording “Total value is copied unchanged from price”, and the proposed user-facing wording “Total value is calculated from unit price and quantity”. Ask the reviewer to explain in plain language why field lineage represents dependencies of an output field on its sources. They should supply an example from an output amount calculated from quantity and price, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: inspect inputFields for each output column and the query expression. A final note should state a many-to-one calculation is not an identity mapping, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 156: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for an output amount calculated from quantity and price. The technical reference output column total_value uses input fields unit_price and quantity must still identify the same object or evidence, while the localized display may read “Total value is calculated from unit price and quantity”. Fail the test if the export instead contains “Total value is copied unchanged from price” where an identifier or protocol value was expected. Run the field-specific assertion: inspect inputFields for each output column and the query expression. This matters because a customer may misunderstand why totals changed. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation a many-to-one calculation is not an identity mapping into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify column lineage output field, but it must retain the source evidence and must pass this field-specific test: inspect inputFields for each output column and the query expression. Transfer destination: a field impact analysis.

Module 27: DIRECT and INDIRECT transformations

In a customer report filtered on active_flag, the source fixture is active_flag contributes INDIRECT FILTER lineage, not the displayed name value. The tempting mistranslation is “The output name contains active_flag values”, whereas the controlled explanation is “The active flag changes which records appear”. This distinction matters because indirect dependency affects inclusion or arrangement without forming output values. A person who does not know the domain can easily overlook that JOIN, SORT, GROUP_BY and WINDOW may also be indirect influences. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 157: Catalog label versus identifier

A reader opens a catalog card for a customer report filtered on active_flag. The underlying evidence is active_flag contributes INDIRECT FILTER lineage, not the displayed name value. The dangerous shortcut is “The output name contains active_flag values”, because it treats a localized display convention as if it were the object itself. A better presentation is “The active flag changes which records appear”. The primary contract is: indirect dependency affects inclusion or arrangement without forming output values. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, compare transformation type and subtype in the referenced facet. The language editor should also note that JOIN, SORT, GROUP_BY and WINDOW may also be indirect influences. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 158: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a customer report filtered on active_flag as ordinary text. It may output “The output name contains active_flag values”, a grammatically plausible change that does not respect the protocol role of active_flag contributes INDIRECT FILTER lineage, not the displayed name value. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “The active flag changes which records appear”. For this field the reason is not stylistic preference: indirect dependency affects inclusion or arrangement without forming output values. A reviewer should apply the concrete test to this example: compare transformation type and subtype in the referenced facet. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that JOIN, SORT, GROUP_BY and WINDOW may also be indirect influences.

Diagnostic 159: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a customer report filtered on active_flag; the resulting translation resembles “The output name contains active_flag values”. Reconstruct the full path and the scenario, a customer report filtered on active_flag, before rewriting the label. Put the original evidence, active_flag contributes INDIRECT FILTER lineage, not the displayed name value, next to the human recommendation, “The active flag changes which records appear”. Then ask whether the source record actually supports the claim: indirect dependency affects inclusion or arrangement without forming output values. A correct diagnosis is reproducible because compare transformation type and subtype in the referenced facet. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: JOIN, SORT, GROUP_BY and WINDOW may also be indirect influences. The goal is reader understanding without inventing a technical fact.

Diagnostic 160: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a customer report filtered on active_flag. If the translation has followed the shortcut “The output name contains active_flag values”, the operational danger is that a privacy review may miss the influence of a sensitive filter field. A better answer begins with the source evidence active_flag contributes INDIRECT FILTER lineage, not the displayed name value, describes it as “The active flag changes which records appear”, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: compare transformation type and subtype in the referenced facet. The caveat JOIN, SORT, GROUP_BY and WINDOW may also be indirect influences also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 161: Bilingual reviewer handoff

Give an independent domain reviewer the source item active_flag contributes INDIRECT FILTER lineage, not the displayed name value, the rejected wording “The output name contains active_flag values”, and the proposed user-facing wording “The active flag changes which records appear”. Ask the reviewer to explain in plain language why indirect dependency affects inclusion or arrangement without forming output values. They should supply an example from a customer report filtered on active_flag, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: compare transformation type and subtype in the referenced facet. A final note should state JOIN, SORT, GROUP_BY and WINDOW may also be indirect influences, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 162: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a customer report filtered on active_flag. The technical reference active_flag contributes INDIRECT FILTER lineage, not the displayed name value must still identify the same object or evidence, while the localized display may read “The active flag changes which records appear”. Fail the test if the export instead contains “The output name contains active_flag values” where an identifier or protocol value was expected. Run the field-specific assertion: compare transformation type and subtype in the referenced facet. This matters because a privacy review may miss the influence of a sensitive filter field. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation JOIN, SORT, GROUP_BY and WINDOW may also be indirect influences into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify direct and indirect transformations, but it must retain the source evidence and must pass this field-specific test: compare transformation type and subtype in the referenced facet. Transfer destination: a graph legend and accessible alt text.

Module 28: Lineage job facet

In a job writing independent customer and order summaries, the source fixture is job.facets.lineage.entries contains two target-specific edge lists. The tempting mistranslation is “Every input feeds every output”, whereas the controlled explanation is “The customer feed produces customer summary; the order feed produces order summary”. This distinction matters because an explicit lineage facet can prevent false Cartesian-product edges. A person who does not know the domain can easily overlook that the official facet may supersede graph edges naively inferred from event boundaries. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 163: Catalog label versus identifier

A reader opens a catalog card for a job writing independent customer and order summaries. The underlying evidence is job.facets.lineage.entries contains two target-specific edge lists. The dangerous shortcut is “Every input feeds every output”, because it treats a localized display convention as if it were the object itself. A better presentation is “The customer feed produces customer summary; the order feed produces order summary”. The primary contract is: an explicit lineage facet can prevent false Cartesian-product edges. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, compare the exact entries and source targets, giving the facet priority as specified. The language editor should also note that the official facet may supersede graph edges naively inferred from event boundaries. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 164: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a job writing independent customer and order summaries as ordinary text. It may output “Every input feeds every output”, a grammatically plausible change that does not respect the protocol role of job.facets.lineage.entries contains two target-specific edge lists. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “The customer feed produces customer summary; the order feed produces order summary”. For this field the reason is not stylistic preference: an explicit lineage facet can prevent false Cartesian-product edges. A reviewer should apply the concrete test to this example: compare the exact entries and source targets, giving the facet priority as specified. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that the official facet may supersede graph edges naively inferred from event boundaries.

Diagnostic 165: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a job writing independent customer and order summaries; the resulting translation resembles “Every input feeds every output”. Reconstruct the full path and the scenario, a job writing independent customer and order summaries, before rewriting the label. Put the original evidence, job.facets.lineage.entries contains two target-specific edge lists, next to the human recommendation, “The customer feed produces customer summary; the order feed produces order summary”. Then ask whether the source record actually supports the claim: an explicit lineage facet can prevent false Cartesian-product edges. A correct diagnosis is reproducible because compare the exact entries and source targets, giving the facet priority as specified. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: the official facet may supersede graph edges naively inferred from event boundaries. The goal is reader understanding without inventing a technical fact.

Diagnostic 166: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a job writing independent customer and order summaries. If the translation has followed the shortcut “Every input feeds every output”, the operational danger is that a translator’s generic “combined inputs” phrase creates imaginary dependencies. A better answer begins with the source evidence job.facets.lineage.entries contains two target-specific edge lists, describes it as “The customer feed produces customer summary; the order feed produces order summary”, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: compare the exact entries and source targets, giving the facet priority as specified. The caveat the official facet may supersede graph edges naively inferred from event boundaries also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 167: Bilingual reviewer handoff

Give an independent domain reviewer the source item job.facets.lineage.entries contains two target-specific edge lists, the rejected wording “Every input feeds every output”, and the proposed user-facing wording “The customer feed produces customer summary; the order feed produces order summary”. Ask the reviewer to explain in plain language why an explicit lineage facet can prevent false Cartesian-product edges. They should supply an example from a job writing independent customer and order summaries, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: compare the exact entries and source targets, giving the facet priority as specified. A final note should state the official facet may supersede graph edges naively inferred from event boundaries, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 168: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a job writing independent customer and order summaries. The technical reference job.facets.lineage.entries contains two target-specific edge lists must still identify the same object or evidence, while the localized display may read “The customer feed produces customer summary; the order feed produces order summary”. Fail the test if the export instead contains “Every input feeds every output” where an identifier or protocol value was expected. Run the field-specific assertion: compare the exact entries and source targets, giving the facet priority as specified. This matters because a translator’s generic “combined inputs” phrase creates imaginary dependencies. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation the official facet may supersede graph edges naively inferred from event boundaries into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify lineage job facet, but it must retain the source evidence and must pass this field-specific test: compare the exact entries and source targets, giving the facet priority as specified. Transfer destination: a two-output pipeline diagram.

Module 29: Lineage dataset facet

In a database view without a natural event job, the source fixture is dataset.facets.lineage.inputs points to raw.customers. The tempting mistranslation is “The view directly executes a parent job”, whereas the controlled explanation is “This cataloged view depends on raw.customers”. This distinction matters because dataset-level lineage may describe structural dependency without a RunEvent owner. A person who does not know the domain can easily overlook that declared/curated lineage should be labeled as such. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 169: Catalog label versus identifier

A reader opens a catalog card for a database view without a natural event job. The underlying evidence is dataset.facets.lineage.inputs points to raw.customers. The dangerous shortcut is “The view directly executes a parent job”, because it treats a localized display convention as if it were the object itself. A better presentation is “This cataloged view depends on raw.customers”. The primary contract is: dataset-level lineage may describe structural dependency without a RunEvent owner. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, inspect DatasetEvent and field map rather than inventing a run history. The language editor should also note that declared/curated lineage should be labeled as such. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 170: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a database view without a natural event job as ordinary text. It may output “The view directly executes a parent job”, a grammatically plausible change that does not respect the protocol role of dataset.facets.lineage.inputs points to raw.customers. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “This cataloged view depends on raw.customers”. For this field the reason is not stylistic preference: dataset-level lineage may describe structural dependency without a RunEvent owner. A reviewer should apply the concrete test to this example: inspect DatasetEvent and field map rather than inventing a run history. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that declared/curated lineage should be labeled as such.

Diagnostic 171: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a database view without a natural event job; the resulting translation resembles “The view directly executes a parent job”. Reconstruct the full path and the scenario, a database view without a natural event job, before rewriting the label. Put the original evidence, dataset.facets.lineage.inputs points to raw.customers, next to the human recommendation, “This cataloged view depends on raw.customers”. Then ask whether the source record actually supports the claim: dataset-level lineage may describe structural dependency without a RunEvent owner. A correct diagnosis is reproducible because inspect DatasetEvent and field map rather than inventing a run history. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: declared/curated lineage should be labeled as such. The goal is reader understanding without inventing a technical fact.

Diagnostic 172: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a database view without a natural event job. If the translation has followed the shortcut “The view directly executes a parent job”, the operational danger is that users may falsely assume observed execution evidence exists. A better answer begins with the source evidence dataset.facets.lineage.inputs points to raw.customers, describes it as “This cataloged view depends on raw.customers”, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: inspect DatasetEvent and field map rather than inventing a run history. The caveat declared/curated lineage should be labeled as such also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 173: Bilingual reviewer handoff

Give an independent domain reviewer the source item dataset.facets.lineage.inputs points to raw.customers, the rejected wording “The view directly executes a parent job”, and the proposed user-facing wording “This cataloged view depends on raw.customers”. Ask the reviewer to explain in plain language why dataset-level lineage may describe structural dependency without a RunEvent owner. They should supply an example from a database view without a natural event job, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: inspect DatasetEvent and field map rather than inventing a run history. A final note should state declared/curated lineage should be labeled as such, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 174: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a database view without a natural event job. The technical reference dataset.facets.lineage.inputs points to raw.customers must still identify the same object or evidence, while the localized display may read “This cataloged view depends on raw.customers”. Fail the test if the export instead contains “The view directly executes a parent job” where an identifier or protocol value was expected. Run the field-specific assertion: inspect DatasetEvent and field map rather than inventing a run history. This matters because users may falsely assume observed execution evidence exists. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation declared/curated lineage should be labeled as such into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify lineage dataset facet, but it must retain the source evidence and must pass this field-specific test: inspect DatasetEvent and field map rather than inventing a run history. Transfer destination: a catalog-only dependency page.

Module 30: Dataset symlink identifiers

In a dataset renamed between catalog migrations, the source fixture is dataset.facets.symlinks.identifiers refers to an alternate name. The tempting mistranslation is “Two unrelated datasets”, whereas the controlled explanation is “Alternate identifier for the same dataset, if documented by the facet”. This distinction matters because symlink identifiers express another reference to an identity, not a second independent lineage edge. A person who does not know the domain can easily overlook that an alias needs disambiguation when names are recycled. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 175: Catalog label versus identifier

A reader opens a catalog card for a dataset renamed between catalog migrations. The underlying evidence is dataset.facets.symlinks.identifiers refers to an alternate name. The dangerous shortcut is “Two unrelated datasets”, because it treats a localized display convention as if it were the object itself. A better presentation is “Alternate identifier for the same dataset, if documented by the facet”. The primary contract is: symlink identifiers express another reference to an identity, not a second independent lineage edge. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, verify authoritative identifier mapping and avoid merging unrelated nodes. The language editor should also note that an alias needs disambiguation when names are recycled. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 176: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a dataset renamed between catalog migrations as ordinary text. It may output “Two unrelated datasets”, a grammatically plausible change that does not respect the protocol role of dataset.facets.symlinks.identifiers refers to an alternate name. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Alternate identifier for the same dataset, if documented by the facet”. For this field the reason is not stylistic preference: symlink identifiers express another reference to an identity, not a second independent lineage edge. A reviewer should apply the concrete test to this example: verify authoritative identifier mapping and avoid merging unrelated nodes. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that an alias needs disambiguation when names are recycled.

Diagnostic 177: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a dataset renamed between catalog migrations; the resulting translation resembles “Two unrelated datasets”. Reconstruct the full path and the scenario, a dataset renamed between catalog migrations, before rewriting the label. Put the original evidence, dataset.facets.symlinks.identifiers refers to an alternate name, next to the human recommendation, “Alternate identifier for the same dataset, if documented by the facet”. Then ask whether the source record actually supports the claim: symlink identifiers express another reference to an identity, not a second independent lineage edge. A correct diagnosis is reproducible because verify authoritative identifier mapping and avoid merging unrelated nodes. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: an alias needs disambiguation when names are recycled. The goal is reader understanding without inventing a technical fact.

Diagnostic 178: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a dataset renamed between catalog migrations. If the translation has followed the shortcut “Two unrelated datasets”, the operational danger is that a translated alias could double-count datasets or erase history. A better answer begins with the source evidence dataset.facets.symlinks.identifiers refers to an alternate name, describes it as “Alternate identifier for the same dataset, if documented by the facet”, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: verify authoritative identifier mapping and avoid merging unrelated nodes. The caveat an alias needs disambiguation when names are recycled also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 179: Bilingual reviewer handoff

Give an independent domain reviewer the source item dataset.facets.symlinks.identifiers refers to an alternate name, the rejected wording “Two unrelated datasets”, and the proposed user-facing wording “Alternate identifier for the same dataset, if documented by the facet”. Ask the reviewer to explain in plain language why symlink identifiers express another reference to an identity, not a second independent lineage edge. They should supply an example from a dataset renamed between catalog migrations, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: verify authoritative identifier mapping and avoid merging unrelated nodes. A final note should state an alias needs disambiguation when names are recycled, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 180: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a dataset renamed between catalog migrations. The technical reference dataset.facets.symlinks.identifiers refers to an alternate name must still identify the same object or evidence, while the localized display may read “Alternate identifier for the same dataset, if documented by the facet”. Fail the test if the export instead contains “Two unrelated datasets” where an identifier or protocol value was expected. Run the field-specific assertion: verify authoritative identifier mapping and avoid merging unrelated nodes. This matters because a translated alias could double-count datasets or erase history. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation an alias needs disambiguation when names are recycled into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify dataset symlink identifiers, but it must retain the source evidence and must pass this field-specific test: verify authoritative identifier mapping and avoid merging unrelated nodes. Transfer destination: a searchable historical catalog.

Ownership, extensions and event interpretation

Module 31: Custom facet name and schema

In a vendor extension carrying regional business labels, the source fixture is custom facets must carry a project prefix and versioned _schemaURL. The tempting mistranslation is “official standard OpenLineage field” assigned to a private extension, whereas the controlled explanation is “Vendor-defined metadata; semantics verified against its referenced schema”. This distinction matters because custom facet keys can have producer-specific contracts and should not collide with standard names. A person who does not know the domain can easily overlook that the display dictionary must record owner, version and provenance. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 181: Catalog label versus identifier

A reader opens a catalog card for a vendor extension carrying regional business labels. The underlying evidence is custom facets must carry a project prefix and versioned _schemaURL. The dangerous shortcut is “official standard OpenLineage field” assigned to a private extension, because it treats a localized display convention as if it were the object itself. A better presentation is “Vendor-defined metadata; semantics verified against its referenced schema”. The primary contract is: custom facet keys can have producer-specific contracts and should not collide with standard names. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, resolve the immutable facet schema and inspect typed fields. The language editor should also note that the display dictionary must record owner, version and provenance. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 182: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a vendor extension carrying regional business labels as ordinary text. It may output “official standard OpenLineage field” assigned to a private extension, a grammatically plausible change that does not respect the protocol role of custom facets must carry a project prefix and versioned _schemaURL. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Vendor-defined metadata; semantics verified against its referenced schema”. For this field the reason is not stylistic preference: custom facet keys can have producer-specific contracts and should not collide with standard names. A reviewer should apply the concrete test to this example: resolve the immutable facet schema and inspect typed fields. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that the display dictionary must record owner, version and provenance.

Diagnostic 183: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a vendor extension carrying regional business labels; the resulting translation resembles “official standard OpenLineage field” assigned to a private extension. Reconstruct the full path and the scenario, a vendor extension carrying regional business labels, before rewriting the label. Put the original evidence, custom facets must carry a project prefix and versioned _schemaURL, next to the human recommendation, “Vendor-defined metadata; semantics verified against its referenced schema”. Then ask whether the source record actually supports the claim: custom facet keys can have producer-specific contracts and should not collide with standard names. A correct diagnosis is reproducible because resolve the immutable facet schema and inspect typed fields. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: the display dictionary must record owner, version and provenance. The goal is reader understanding without inventing a technical fact.

Diagnostic 184: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a vendor extension carrying regional business labels. If the translation has followed the shortcut “official standard OpenLineage field” assigned to a private extension, the operational danger is that a translator may attach a false standard meaning to custom fields. A better answer begins with the source evidence custom facets must carry a project prefix and versioned _schemaURL, describes it as “Vendor-defined metadata; semantics verified against its referenced schema”, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: resolve the immutable facet schema and inspect typed fields. The caveat the display dictionary must record owner, version and provenance also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 185: Bilingual reviewer handoff

Give an independent domain reviewer the source item custom facets must carry a project prefix and versioned _schemaURL, the rejected wording “official standard OpenLineage field” assigned to a private extension, and the proposed user-facing wording “Vendor-defined metadata; semantics verified against its referenced schema”. Ask the reviewer to explain in plain language why custom facet keys can have producer-specific contracts and should not collide with standard names. They should supply an example from a vendor extension carrying regional business labels, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: resolve the immutable facet schema and inspect typed fields. A final note should state the display dictionary must record owner, version and provenance, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 186: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a vendor extension carrying regional business labels. The technical reference custom facets must carry a project prefix and versioned _schemaURL must still identify the same object or evidence, while the localized display may read “Vendor-defined metadata; semantics verified against its referenced schema”. Fail the test if the export instead contains “official standard OpenLineage field” assigned to a private extension where an identifier or protocol value was expected. Run the field-specific assertion: resolve the immutable facet schema and inspect typed fields. This matters because a translator may attach a false standard meaning to custom fields. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation the display dictionary must record owner, version and provenance into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify custom facet name and schema, but it must retain the source evidence and must pass this field-specific test: resolve the immutable facet schema and inspect typed fields. Transfer destination: third-party connector localization.

Module 32: Ownership job facet

In a shared transformation maintained by a data platform team, the source fixture is job.facets.ownership.owners identifies accountable parties. The tempting mistranslation is “event produced by the team that owns the data”, whereas the controlled explanation is “Job owner” distinguished from metadata producer and dataset steward. This distinction matters because ownership roles and producer identity solve different governance questions. A person who does not know the domain can easily overlook that human names may be transliterated while identity links stay fixed. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 187: Catalog label versus identifier

A reader opens a catalog card for a shared transformation maintained by a data platform team. The underlying evidence is job.facets.ownership.owners identifies accountable parties. The dangerous shortcut is “event produced by the team that owns the data”, because it treats a localized display convention as if it were the object itself. A better presentation is “Job owner” distinguished from metadata producer and dataset steward. The primary contract is: ownership roles and producer identity solve different governance questions. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, compare facet record with operational responsibility directory. The language editor should also note that human names may be transliterated while identity links stay fixed. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 188: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a shared transformation maintained by a data platform team as ordinary text. It may output “event produced by the team that owns the data”, a grammatically plausible change that does not respect the protocol role of job.facets.ownership.owners identifies accountable parties. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Job owner” distinguished from metadata producer and dataset steward. For this field the reason is not stylistic preference: ownership roles and producer identity solve different governance questions. A reviewer should apply the concrete test to this example: compare facet record with operational responsibility directory. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that human names may be transliterated while identity links stay fixed.

Diagnostic 189: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a shared transformation maintained by a data platform team; the resulting translation resembles “event produced by the team that owns the data”. Reconstruct the full path and the scenario, a shared transformation maintained by a data platform team, before rewriting the label. Put the original evidence, job.facets.ownership.owners identifies accountable parties, next to the human recommendation, “Job owner” distinguished from metadata producer and dataset steward. Then ask whether the source record actually supports the claim: ownership roles and producer identity solve different governance questions. A correct diagnosis is reproducible because compare facet record with operational responsibility directory. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: human names may be transliterated while identity links stay fixed. The goal is reader understanding without inventing a technical fact.

Diagnostic 190: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a shared transformation maintained by a data platform team. If the translation has followed the shortcut “event produced by the team that owns the data”, the operational danger is that an incident notification may reach a team that cannot fix the problem. A better answer begins with the source evidence job.facets.ownership.owners identifies accountable parties, describes it as “Job owner” distinguished from metadata producer and dataset steward, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: compare facet record with operational responsibility directory. The caveat human names may be transliterated while identity links stay fixed also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 191: Bilingual reviewer handoff

Give an independent domain reviewer the source item job.facets.ownership.owners identifies accountable parties, the rejected wording “event produced by the team that owns the data”, and the proposed user-facing wording “Job owner” distinguished from metadata producer and dataset steward. Ask the reviewer to explain in plain language why ownership roles and producer identity solve different governance questions. They should supply an example from a shared transformation maintained by a data platform team, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: compare facet record with operational responsibility directory. A final note should state human names may be transliterated while identity links stay fixed, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 192: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a shared transformation maintained by a data platform team. The technical reference job.facets.ownership.owners identifies accountable parties must still identify the same object or evidence, while the localized display may read “Job owner” distinguished from metadata producer and dataset steward. Fail the test if the export instead contains “event produced by the team that owns the data” where an identifier or protocol value was expected. Run the field-specific assertion: compare facet record with operational responsibility directory. This matters because an incident notification may reach a team that cannot fix the problem. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation human names may be transliterated while identity links stay fixed into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify ownership job facet, but it must retain the source evidence and must pass this field-specific test: compare facet record with operational responsibility directory. Transfer destination: an on-call escalation card.

Module 33: Job tags and business labels

In a pipeline tagged finance, critical and daily, the source fixture is job.facets.tags.tags contains key/value metadata. The tempting mistranslation is tag values silently translated into new filter keys, whereas the controlled explanation is “Finance · critical · daily” in a display glossary when approved. This distinction matters because whether tags are controlled keys or display labels depends on their producer contract. A person who does not know the domain can easily overlook that some tag values are programmatic while others are documentation. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 193: Catalog label versus identifier

A reader opens a catalog card for a pipeline tagged finance, critical and daily. The underlying evidence is job.facets.tags.tags contains key/value metadata. The dangerous shortcut is tag values silently translated into new filter keys, because it treats a localized display convention as if it were the object itself. A better presentation is “Finance · critical · daily” in a display glossary when approved. The primary contract is: whether tags are controlled keys or display labels depends on their producer contract. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, test filter results for both language settings against the same job IDs. The language editor should also note that some tag values are programmatic while others are documentation. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 194: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a pipeline tagged finance, critical and daily as ordinary text. It may output tag values silently translated into new filter keys, a grammatically plausible change that does not respect the protocol role of job.facets.tags.tags contains key/value metadata. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Finance · critical · daily” in a display glossary when approved. For this field the reason is not stylistic preference: whether tags are controlled keys or display labels depends on their producer contract. A reviewer should apply the concrete test to this example: test filter results for both language settings against the same job IDs. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that some tag values are programmatic while others are documentation.

Diagnostic 195: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a pipeline tagged finance, critical and daily; the resulting translation resembles tag values silently translated into new filter keys. Reconstruct the full path and the scenario, a pipeline tagged finance, critical and daily, before rewriting the label. Put the original evidence, job.facets.tags.tags contains key/value metadata, next to the human recommendation, “Finance · critical · daily” in a display glossary when approved. Then ask whether the source record actually supports the claim: whether tags are controlled keys or display labels depends on their producer contract. A correct diagnosis is reproducible because test filter results for both language settings against the same job IDs. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: some tag values are programmatic while others are documentation. The goal is reader understanding without inventing a technical fact.

Diagnostic 196: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a pipeline tagged finance, critical and daily. If the translation has followed the shortcut tag values silently translated into new filter keys, the operational danger is that localized filtering may hide production incidents. A better answer begins with the source evidence job.facets.tags.tags contains key/value metadata, describes it as “Finance · critical · daily” in a display glossary when approved, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: test filter results for both language settings against the same job IDs. The caveat some tag values are programmatic while others are documentation also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 197: Bilingual reviewer handoff

Give an independent domain reviewer the source item job.facets.tags.tags contains key/value metadata, the rejected wording tag values silently translated into new filter keys, and the proposed user-facing wording “Finance · critical · daily” in a display glossary when approved. Ask the reviewer to explain in plain language why whether tags are controlled keys or display labels depends on their producer contract. They should supply an example from a pipeline tagged finance, critical and daily, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: test filter results for both language settings against the same job IDs. A final note should state some tag values are programmatic while others are documentation, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 198: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a pipeline tagged finance, critical and daily. The technical reference job.facets.tags.tags contains key/value metadata must still identify the same object or evidence, while the localized display may read “Finance · critical · daily” in a display glossary when approved. Fail the test if the export instead contains tag values silently translated into new filter keys where an identifier or protocol value was expected. Run the field-specific assertion: test filter results for both language settings against the same job IDs. This matters because localized filtering may hide production incidents. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation some tag values are programmatic while others are documentation into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify job tags and business labels, but it must retain the source evidence and must pass this field-specific test: test filter results for both language settings against the same job IDs. Transfer destination: search facets on an operations site.

Module 34: Source code and location references

In a job associated with a versioned repository path, the source fixture is sourceCodeLocation-like facet points to a repository URL and commit. The tempting mistranslation is repository path translated as a sentence, whereas the controlled explanation is “View source for this pipeline” as a separate button label. This distinction matters because code location and commit reference are evidentiary pointers, not editable text. A person who does not know the domain can easily overlook that do not assume every integration emits the same source-code facet shape. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 199: Catalog label versus identifier

A reader opens a catalog card for a job associated with a versioned repository path. The underlying evidence is sourceCodeLocation-like facet points to a repository URL and commit. The dangerous shortcut is repository path translated as a sentence, because it treats a localized display convention as if it were the object itself. A better presentation is “View source for this pipeline” as a separate button label. The primary contract is: code location and commit reference are evidentiary pointers, not editable text. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, compare exact URL and commit/hash in localized view. The language editor should also note that do not assume every integration emits the same source-code facet shape. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 200: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a job associated with a versioned repository path as ordinary text. It may output repository path translated as a sentence, a grammatically plausible change that does not respect the protocol role of sourceCodeLocation-like facet points to a repository URL and commit. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “View source for this pipeline” as a separate button label. For this field the reason is not stylistic preference: code location and commit reference are evidentiary pointers, not editable text. A reviewer should apply the concrete test to this example: compare exact URL and commit/hash in localized view. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that do not assume every integration emits the same source-code facet shape.

Diagnostic 201: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a job associated with a versioned repository path; the resulting translation resembles repository path translated as a sentence. Reconstruct the full path and the scenario, a job associated with a versioned repository path, before rewriting the label. Put the original evidence, sourceCodeLocation-like facet points to a repository URL and commit, next to the human recommendation, “View source for this pipeline” as a separate button label. Then ask whether the source record actually supports the claim: code location and commit reference are evidentiary pointers, not editable text. A correct diagnosis is reproducible because compare exact URL and commit/hash in localized view. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: do not assume every integration emits the same source-code facet shape. The goal is reader understanding without inventing a technical fact.

Diagnostic 202: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a job associated with a versioned repository path. If the translation has followed the shortcut repository path translated as a sentence, the operational danger is that a reviewer may inspect the wrong code revision. A better answer begins with the source evidence sourceCodeLocation-like facet points to a repository URL and commit, describes it as “View source for this pipeline” as a separate button label, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: compare exact URL and commit/hash in localized view. The caveat do not assume every integration emits the same source-code facet shape also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 203: Bilingual reviewer handoff

Give an independent domain reviewer the source item sourceCodeLocation-like facet points to a repository URL and commit, the rejected wording repository path translated as a sentence, and the proposed user-facing wording “View source for this pipeline” as a separate button label. Ask the reviewer to explain in plain language why code location and commit reference are evidentiary pointers, not editable text. They should supply an example from a job associated with a versioned repository path, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: compare exact URL and commit/hash in localized view. A final note should state do not assume every integration emits the same source-code facet shape, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 204: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a job associated with a versioned repository path. The technical reference sourceCodeLocation-like facet points to a repository URL and commit must still identify the same object or evidence, while the localized display may read “View source for this pipeline” as a separate button label. Fail the test if the export instead contains repository path translated as a sentence where an identifier or protocol value was expected. Run the field-specific assertion: compare exact URL and commit/hash in localized view. This matters because a reviewer may inspect the wrong code revision. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation do not assume every integration emits the same source-code facet shape into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify source code and location references, but it must retain the source evidence and must pass this field-specific test: compare exact URL and commit/hash in localized view. Transfer destination: reproducible code review.

Module 35: Dataset type facet

In a materialized analytics view, the source fixture is dataset.facets.datasetType.datasetType = VIEW; subType = MATERIALIZED. The tempting mistranslation is “Source table” presented to readers, whereas the controlled explanation is “Materialized view” with exact type codes retained. This distinction matters because technical storage type affects how a dataset is refreshed and queried. A person who does not know the domain can easily overlook that type and subtype deserve separate labels. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 205: Catalog label versus identifier

A reader opens a catalog card for a materialized analytics view. The underlying evidence is dataset.facets.datasetType.datasetType = VIEW; subType = MATERIALIZED. The dangerous shortcut is “Source table” presented to readers, because it treats a localized display convention as if it were the object itself. A better presentation is “Materialized view” with exact type codes retained. The primary contract is: technical storage type affects how a dataset is refreshed and queried. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, read datasetType and subType rather than infer from UI icons. The language editor should also note that type and subtype deserve separate labels. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 206: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a materialized analytics view as ordinary text. It may output “Source table” presented to readers, a grammatically plausible change that does not respect the protocol role of dataset.facets.datasetType.datasetType = VIEW; subType = MATERIALIZED. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Materialized view” with exact type codes retained. For this field the reason is not stylistic preference: technical storage type affects how a dataset is refreshed and queried. A reviewer should apply the concrete test to this example: read datasetType and subType rather than infer from UI icons. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that type and subtype deserve separate labels.

Diagnostic 207: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a materialized analytics view; the resulting translation resembles “Source table” presented to readers. Reconstruct the full path and the scenario, a materialized analytics view, before rewriting the label. Put the original evidence, dataset.facets.datasetType.datasetType = VIEW; subType = MATERIALIZED, next to the human recommendation, “Materialized view” with exact type codes retained. Then ask whether the source record actually supports the claim: technical storage type affects how a dataset is refreshed and queried. A correct diagnosis is reproducible because read datasetType and subType rather than infer from UI icons. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: type and subtype deserve separate labels. The goal is reader understanding without inventing a technical fact.

Diagnostic 208: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a materialized analytics view. If the translation has followed the shortcut “Source table” presented to readers, the operational danger is that operations teams may expect updates that a static table would receive. A better answer begins with the source evidence dataset.facets.datasetType.datasetType = VIEW; subType = MATERIALIZED, describes it as “Materialized view” with exact type codes retained, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: read datasetType and subType rather than infer from UI icons. The caveat type and subtype deserve separate labels also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 209: Bilingual reviewer handoff

Give an independent domain reviewer the source item dataset.facets.datasetType.datasetType = VIEW; subType = MATERIALIZED, the rejected wording “Source table” presented to readers, and the proposed user-facing wording “Materialized view” with exact type codes retained. Ask the reviewer to explain in plain language why technical storage type affects how a dataset is refreshed and queried. They should supply an example from a materialized analytics view, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: read datasetType and subType rather than infer from UI icons. A final note should state type and subtype deserve separate labels, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 210: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a materialized analytics view. The technical reference dataset.facets.datasetType.datasetType = VIEW; subType = MATERIALIZED must still identify the same object or evidence, while the localized display may read “Materialized view” with exact type codes retained. Fail the test if the export instead contains “Source table” presented to readers where an identifier or protocol value was expected. Run the field-specific assertion: read datasetType and subType rather than infer from UI icons. This matters because operations teams may expect updates that a static table would receive. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation type and subtype deserve separate labels into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify dataset type facet, but it must retain the source evidence and must pass this field-specific test: read datasetType and subType rather than infer from UI icons. Transfer destination: a catalog dataset summary.

Module 36: Input/output facet placement

In a data job with input quality and output volume facets, the source fixture is inputs[0].inputFacets and outputs[0].outputFacets have distinct scope. The tempting mistranslation is one generic “quality data” property applied to both sides, whereas the controlled explanation is “Input observation” separated from “output write statistics”. This distinction matters because facet placement communicates whether data was read, produced or measured. A person who does not know the domain can easily overlook that common dataset facets under facets have a different scope from inputFacets or outputFacets. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 211: Catalog label versus identifier

A reader opens a catalog card for a data job with input quality and output volume facets. The underlying evidence is inputs[0].inputFacets and outputs[0].outputFacets have distinct scope. The dangerous shortcut is one generic “quality data” property applied to both sides, because it treats a localized display convention as if it were the object itself. A better presentation is “Input observation” separated from “output write statistics”. The primary contract is: facet placement communicates whether data was read, produced or measured. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, compare object paths, not just field names, in test snapshots. The language editor should also note that common dataset facets under facets have a different scope from inputFacets or outputFacets. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 212: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a data job with input quality and output volume facets as ordinary text. It may output one generic “quality data” property applied to both sides, a grammatically plausible change that does not respect the protocol role of inputs[0].inputFacets and outputs[0].outputFacets have distinct scope. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Input observation” separated from “output write statistics”. For this field the reason is not stylistic preference: facet placement communicates whether data was read, produced or measured. A reviewer should apply the concrete test to this example: compare object paths, not just field names, in test snapshots. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that common dataset facets under facets have a different scope from inputFacets or outputFacets.

Diagnostic 213: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a data job with input quality and output volume facets; the resulting translation resembles one generic “quality data” property applied to both sides. Reconstruct the full path and the scenario, a data job with input quality and output volume facets, before rewriting the label. Put the original evidence, inputs[0].inputFacets and outputs[0].outputFacets have distinct scope, next to the human recommendation, “Input observation” separated from “output write statistics”. Then ask whether the source record actually supports the claim: facet placement communicates whether data was read, produced or measured. A correct diagnosis is reproducible because compare object paths, not just field names, in test snapshots. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: common dataset facets under facets have a different scope from inputFacets or outputFacets. The goal is reader understanding without inventing a technical fact.

Diagnostic 214: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a data job with input quality and output volume facets. If the translation has followed the shortcut one generic “quality data” property applied to both sides, the operational danger is that an analyst may attribute an input defect to the generated output. A better answer begins with the source evidence inputs[0].inputFacets and outputs[0].outputFacets have distinct scope, describes it as “Input observation” separated from “output write statistics”, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: compare object paths, not just field names, in test snapshots. The caveat common dataset facets under facets have a different scope from inputFacets or outputFacets also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 215: Bilingual reviewer handoff

Give an independent domain reviewer the source item inputs[0].inputFacets and outputs[0].outputFacets have distinct scope, the rejected wording one generic “quality data” property applied to both sides, and the proposed user-facing wording “Input observation” separated from “output write statistics”. Ask the reviewer to explain in plain language why facet placement communicates whether data was read, produced or measured. They should supply an example from a data job with input quality and output volume facets, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: compare object paths, not just field names, in test snapshots. A final note should state common dataset facets under facets have a different scope from inputFacets or outputFacets, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 216: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a data job with input quality and output volume facets. The technical reference inputs[0].inputFacets and outputs[0].outputFacets have distinct scope must still identify the same object or evidence, while the localized display may read “Input observation” separated from “output write statistics”. Fail the test if the export instead contains one generic “quality data” property applied to both sides where an identifier or protocol value was expected. Run the field-specific assertion: compare object paths, not just field names, in test snapshots. This matters because an analyst may attribute an input defect to the generated output. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation common dataset facets under facets have a different scope from inputFacets or outputFacets into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify input/output facet placement, but it must retain the source evidence and must pass this field-specific test: compare object paths, not just field names, in test snapshots. Transfer destination: a side-by-side data-quality view.

Module 37: JobEvent versus RunEvent

In a catalog declaring a job before execution, the source fixture is JobEvent records a Job, while RunEvent describes an execution. The tempting mistranslation is “This job succeeded” based only on static JobEvent metadata, whereas the controlled explanation is “Job definition exists; no run result is implied”. This distinction matters because event classes reflect different evidence levels and require distinct claims. A person who does not know the domain can easily overlook that the event schema has distinct object forms that cannot be flattened to prose. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 217: Catalog label versus identifier

A reader opens a catalog card for a catalog declaring a job before execution. The underlying evidence is JobEvent records a Job, while RunEvent describes an execution. The dangerous shortcut is “This job succeeded” based only on static JobEvent metadata, because it treats a localized display convention as if it were the object itself. A better presentation is “Job definition exists; no run result is implied”. The primary contract is: event classes reflect different evidence levels and require distinct claims. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, look for matching runId and terminal state before reporting execution. The language editor should also note that the event schema has distinct object forms that cannot be flattened to prose. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 218: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a catalog declaring a job before execution as ordinary text. It may output “This job succeeded” based only on static JobEvent metadata, a grammatically plausible change that does not respect the protocol role of JobEvent records a Job, while RunEvent describes an execution. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “Job definition exists; no run result is implied”. For this field the reason is not stylistic preference: event classes reflect different evidence levels and require distinct claims. A reviewer should apply the concrete test to this example: look for matching runId and terminal state before reporting execution. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that the event schema has distinct object forms that cannot be flattened to prose.

Diagnostic 219: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a catalog declaring a job before execution; the resulting translation resembles “This job succeeded” based only on static JobEvent metadata. Reconstruct the full path and the scenario, a catalog declaring a job before execution, before rewriting the label. Put the original evidence, JobEvent records a Job, while RunEvent describes an execution, next to the human recommendation, “Job definition exists; no run result is implied”. Then ask whether the source record actually supports the claim: event classes reflect different evidence levels and require distinct claims. A correct diagnosis is reproducible because look for matching runId and terminal state before reporting execution. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: the event schema has distinct object forms that cannot be flattened to prose. The goal is reader understanding without inventing a technical fact.

Diagnostic 220: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a catalog declaring a job before execution. If the translation has followed the shortcut “This job succeeded” based only on static JobEvent metadata, the operational danger is that a status board could invent a successful run. A better answer begins with the source evidence JobEvent records a Job, while RunEvent describes an execution, describes it as “Job definition exists; no run result is implied”, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: look for matching runId and terminal state before reporting execution. The caveat the event schema has distinct object forms that cannot be flattened to prose also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 221: Bilingual reviewer handoff

Give an independent domain reviewer the source item JobEvent records a Job, while RunEvent describes an execution, the rejected wording “This job succeeded” based only on static JobEvent metadata, and the proposed user-facing wording “Job definition exists; no run result is implied”. Ask the reviewer to explain in plain language why event classes reflect different evidence levels and require distinct claims. They should supply an example from a catalog declaring a job before execution, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: look for matching runId and terminal state before reporting execution. A final note should state the event schema has distinct object forms that cannot be flattened to prose, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 222: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a catalog declaring a job before execution. The technical reference JobEvent records a Job, while RunEvent describes an execution must still identify the same object or evidence, while the localized display may read “Job definition exists; no run result is implied”. Fail the test if the export instead contains “This job succeeded” based only on static JobEvent metadata where an identifier or protocol value was expected. Run the field-specific assertion: look for matching runId and terminal state before reporting execution. This matters because a status board could invent a successful run. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation the event schema has distinct object forms that cannot be flattened to prose into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify jobevent versus runevent, but it must retain the source evidence and must pass this field-specific test: look for matching runId and terminal state before reporting execution. Transfer destination: a catalog launch announcement.

Module 38: DatasetEvent versus RunEvent

In a catalog publishing lineage for a static database view, the source fixture is DatasetEvent expresses dataset metadata without claiming a run. The tempting mistranslation is “This view was rebuilt today” inferred from DatasetEvent time, whereas the controlled explanation is “The dataset metadata record was published today”. This distinction matters because a dataset event’s creation time and a data refresh time are not equivalent. A person who does not know the domain can easily overlook that a structural dependency can be valid without a corresponding event job. The learning objective is to describe the field without changing its referent, type, scope, relationship or provenance. The six diagnostics below probe the same field under different real editorial pressures, so the reader can practise making a decision instead of simply memorizing a translation.

Diagnostic 223: Catalog label versus identifier

A reader opens a catalog card for a catalog publishing lineage for a static database view. The underlying evidence is DatasetEvent expresses dataset metadata without claiming a run. The dangerous shortcut is “This view was rebuilt today” inferred from DatasetEvent time, because it treats a localized display convention as if it were the object itself. A better presentation is “The dataset metadata record was published today”. The primary contract is: a dataset event’s creation time and a data refresh time are not equivalent. The card should show the exact technical reference in a detail view even when the main heading is translated. Before approval, inspect dataset version, underlying store and any relevant RunEvent. The language editor should also note that a structural dependency can be valid without a corresponding event job. This is a small change on screen but a large difference in what a reader is entitled to infer.

Diagnostic 224: Machine-translation trap and bounded repair

Suppose a translation engine receives the whole JSON fragment for a catalog publishing lineage for a static database view as ordinary text. It may output “This view was rebuilt today” inferred from DatasetEvent time, a grammatically plausible change that does not respect the protocol role of DatasetEvent expresses dataset metadata without claiming a run. The translation brief must whitelist only the human-visible explanation and explicitly protect identifiers, enum values, paths, types, timestamps and machine schema references. The approved rendering instead says “The dataset metadata record was published today”. For this field the reason is not stylistic preference: a dataset event’s creation time and a data refresh time are not equivalent. A reviewer should apply the concrete test to this example: inspect dataset version, underlying store and any relevant RunEvent. If the system cannot distinguish prose from data, pause automatic translation for this item and use a reviewed glossary entry keyed by the original path. Add a note that a structural dependency can be valid without a corresponding event job.

Diagnostic 225: Conflicting context and semantic diagnosis

Now place the same term in a second screen that omits the surrounding technical context. The editor sees only a short label and may infer the wrong subject of a catalog publishing lineage for a static database view; the resulting translation resembles “This view was rebuilt today” inferred from DatasetEvent time. Reconstruct the full path and the scenario, a catalog publishing lineage for a static database view, before rewriting the label. Put the original evidence, DatasetEvent expresses dataset metadata without claiming a run, next to the human recommendation, “The dataset metadata record was published today”. Then ask whether the source record actually supports the claim: a dataset event’s creation time and a data refresh time are not equivalent. A correct diagnosis is reproducible because inspect dataset version, underlying store and any relevant RunEvent. When a reviewer is uncertain, preserve the precise source term and an explanatory note rather than smoothing away the ambiguity. Remember the additional complication: a structural dependency can be valid without a corresponding event job. The goal is reader understanding without inventing a technical fact.

Diagnostic 226: Incident question and consequence

During an investigation, a colleague asks whether the localized field explains an unexpected result in a catalog publishing lineage for a static database view. If the translation has followed the shortcut “This view was rebuilt today” inferred from DatasetEvent time, the operational danger is that freshness claims may be overstated. A better answer begins with the source evidence DatasetEvent expresses dataset metadata without claiming a run, describes it as “The dataset metadata record was published today”, and avoids claims beyond its actual scope. The investigator can establish truth by doing this specific check: inspect dataset version, underlying store and any relevant RunEvent. The caveat a structural dependency can be valid without a corresponding event job also belongs in the incident notes whenever it could change interpretation. This case shows why a translator must be able to say “not established by this field” rather than turning an incomplete fact into a reassuring, fluent sentence. A bilingual incident report should preserve the same uncertainty in both versions.

Diagnostic 227: Bilingual reviewer handoff

Give an independent domain reviewer the source item DatasetEvent expresses dataset metadata without claiming a run, the rejected wording “This view was rebuilt today” inferred from DatasetEvent time, and the proposed user-facing wording “The dataset metadata record was published today”. Ask the reviewer to explain in plain language why a dataset event’s creation time and a data refresh time are not equivalent. They should supply an example from a catalog publishing lineage for a static database view, not merely approve punctuation or tone. The translation record then contains four separate columns: original technical value, target-language display text, verification evidence and unresolved cautions. For verification use the specific procedure: inspect dataset version, underlying store and any relevant RunEvent. A final note should state a structural dependency can be valid without a corresponding event job, because a second translator may otherwise reuse this phrase where the conditions differ. The handoff is complete only when a different reader can reproduce the same interpretation without consulting the author of the first translation.

Diagnostic 228: Release regression and transfer test

Switch the interface from the source language to the target language and back while keeping the same fixture for a catalog publishing lineage for a static database view. The technical reference DatasetEvent expresses dataset metadata without claiming a run must still identify the same object or evidence, while the localized display may read “The dataset metadata record was published today”. Fail the test if the export instead contains “This view was rebuilt today” inferred from DatasetEvent time where an identifier or protocol value was expected. Run the field-specific assertion: inspect dataset version, underlying store and any relevant RunEvent. This matters because freshness claims may be overstated. Inspect search, deep links, graph edges, keyboard navigation and any exported report as applicable. Carry the limitation a structural dependency can be valid without a corresponding event job into the release note. The result should be a reversible translation-layer change, not a silent mutation to the pipeline history or a false additional record.

Module acceptance rule: The human translation may clarify datasetevent versus runevent, but it must retain the source evidence and must pass this field-specific test: inspect dataset version, underlying store and any relevant RunEvent. Transfer destination: a data discovery page.

Transfer: end-to-end workshops for practitioners

Workshop 1: A retail orders pipeline

Represent purchases as an input and the daily order summary as an output. Localize the reader’s interpretation of the two datasets, then compare the graph against the source event. Verify that the translation says whether totals include cancelled orders, tax or refunds only when the source rules prove it. Begin with a fictional read-only event fixture and make a compact source-to-target dictionary containing the exact field path, raw value, explanation, language tag, reviewer and evidence reference. Have one person who did not write the text trace the same object from the localized screen to the raw event. Reject the draft if that person encounters a second identity, a changed time, an unsupported business assertion or an unexplained edge.

For the final comparison, save a technical diff of the pre-translation and post-translation exports. Classify each changed byte: display-only prose, approved locale formatting, documented alias, or prohibited protocol mutation. Explain in writing why every accepted change belongs to its class. The workshop passes when the target-language reading is clearer and the engineering reviewer can still independently reconstruct the original run, job, dataset and dependencies. Reuse this audit method for a different domain instead of relying on a memorized phrase.

Workshop 2: A multilingual customer support dashboard

The display title uses a translated business noun while its dataset namespace and name remain unchanged. Check that search facets, copied deep links, incident notices and audit exports resolve to one consistent technical record for every locale. Begin with a fictional read-only event fixture and make a compact source-to-target dictionary containing the exact field path, raw value, explanation, language tag, reviewer and evidence reference. Have one person who did not write the text trace the same object from the localized screen to the raw event. Reject the draft if that person encounters a second identity, a changed time, an unsupported business assertion or an unexplained edge.

For the final comparison, save a technical diff of the pre-translation and post-translation exports. Classify each changed byte: display-only prose, approved locale formatting, documented alias, or prohibited protocol mutation. Explain in writing why every accepted change belongs to its class. The workshop passes when the target-language reading is clearer and the engineering reviewer can still independently reconstruct the original run, job, dataset and dependencies. Reuse this audit method for a different domain instead of relying on a memorized phrase.

Workshop 3: A finance settlement incident

Compare a FAIL event with the business date of the settlement report. Test whether the translated incident timeline distinguishes failed execution, nominal period and historical business values. Preserve exact runId and owner routing. Begin with a fictional read-only event fixture and make a compact source-to-target dictionary containing the exact field path, raw value, explanation, language tag, reviewer and evidence reference. Have one person who did not write the text trace the same object from the localized screen to the raw event. Reject the draft if that person encounters a second identity, a changed time, an unsupported business assertion or an unexplained edge.

For the final comparison, save a technical diff of the pre-translation and post-translation exports. Classify each changed byte: display-only prose, approved locale formatting, documented alias, or prohibited protocol mutation. Explain in writing why every accepted change belongs to its class. The workshop passes when the target-language reading is clearer and the engineering reviewer can still independently reconstruct the original run, job, dataset and dependencies. Reuse this audit method for a different domain instead of relying on a memorized phrase.

Workshop 4: A data quality alert

Translate a nullCount metric into a precise sentence that names the field, population, measurement time and count. Ask whether null really implies invalidity. Reject a version that claims overall data integrity from one narrowly scoped assertion. Begin with a fictional read-only event fixture and make a compact source-to-target dictionary containing the exact field path, raw value, explanation, language tag, reviewer and evidence reference. Have one person who did not write the text trace the same object from the localized screen to the raw event. Reject the draft if that person encounters a second identity, a changed time, an unsupported business assertion or an unexplained edge.

For the final comparison, save a technical diff of the pre-translation and post-translation exports. Classify each changed byte: display-only prose, approved locale formatting, documented alias, or prohibited protocol mutation. Explain in writing why every accepted change belongs to its class. The workshop passes when the target-language reading is clearer and the engineering reviewer can still independently reconstruct the original run, job, dataset and dependencies. Reuse this audit method for a different domain instead of relying on a memorized phrase.

Workshop 5: A table-to-view migration

Prepare two descriptions: the old source table and the new materialized view. Check lifecycle changes, aliases and dataset versions. Do not let a friendly localized label imply that the migration happened without recorded evidence. Begin with a fictional read-only event fixture and make a compact source-to-target dictionary containing the exact field path, raw value, explanation, language tag, reviewer and evidence reference. Have one person who did not write the text trace the same object from the localized screen to the raw event. Reject the draft if that person encounters a second identity, a changed time, an unsupported business assertion or an unexplained edge.

For the final comparison, save a technical diff of the pre-translation and post-translation exports. Classify each changed byte: display-only prose, approved locale formatting, documented alias, or prohibited protocol mutation. Explain in writing why every accepted change belongs to its class. The workshop passes when the target-language reading is clearer and the engineering reviewer can still independently reconstruct the original run, job, dataset and dependencies. Reuse this audit method for a different domain instead of relying on a memorized phrase.

Workshop 6: A nested field schema

Construct a dataset with a nested address object, an optional postal code and a numeric customer ID. Produce localized descriptions for each field while keeping paths and types. Test that the same nested field appears in an export. Begin with a fictional read-only event fixture and make a compact source-to-target dictionary containing the exact field path, raw value, explanation, language tag, reviewer and evidence reference. Have one person who did not write the text trace the same object from the localized screen to the raw event. Reject the draft if that person encounters a second identity, a changed time, an unsupported business assertion or an unexplained edge.

For the final comparison, save a technical diff of the pre-translation and post-translation exports. Classify each changed byte: display-only prose, approved locale formatting, documented alias, or prohibited protocol mutation. Explain in writing why every accepted change belongs to its class. The workshop passes when the target-language reading is clearer and the engineering reviewer can still independently reconstruct the original run, job, dataset and dependencies. Reuse this audit method for a different domain instead of relying on a memorized phrase.

Workshop 7: An SQL investigation

Explain a filtering query in ordinary language without translating SQL syntax or string literals. Use column lineage to identify direct and indirect influences and explain why one source field determines which output rows survive. Begin with a fictional read-only event fixture and make a compact source-to-target dictionary containing the exact field path, raw value, explanation, language tag, reviewer and evidence reference. Have one person who did not write the text trace the same object from the localized screen to the raw event. Reject the draft if that person encounters a second identity, a changed time, an unsupported business assertion or an unexplained edge.

For the final comparison, save a technical diff of the pre-translation and post-translation exports. Classify each changed byte: display-only prose, approved locale formatting, documented alias, or prohibited protocol mutation. Explain in writing why every accepted change belongs to its class. The workshop passes when the target-language reading is clearer and the engineering reviewer can still independently reconstruct the original run, job, dataset and dependencies. Reuse this audit method for a different domain instead of relying on a memorized phrase.

Workshop 8: An orchestrator with child tasks

Link three child runs to a parent execution while describing each step in the target language. Verify that the translation does not assert additional data dependency or ownership merely because the parent relation exists. Begin with a fictional read-only event fixture and make a compact source-to-target dictionary containing the exact field path, raw value, explanation, language tag, reviewer and evidence reference. Have one person who did not write the text trace the same object from the localized screen to the raw event. Reject the draft if that person encounters a second identity, a changed time, an unsupported business assertion or an unexplained edge.

For the final comparison, save a technical diff of the pre-translation and post-translation exports. Classify each changed byte: display-only prose, approved locale formatting, documented alias, or prohibited protocol mutation. Explain in writing why every accepted change belongs to its class. The workshop passes when the target-language reading is clearer and the engineering reviewer can still independently reconstruct the original run, job, dataset and dependencies. Reuse this audit method for a different domain instead of relying on a memorized phrase.

Workshop 9: A streaming ingestion service

Represent START, RUNNING, OTHER and terminal states accurately. Make a display mapping for progress messages without turning a nonterminal metadata update into “complete”. Verify what the stream processor actually emitted. Begin with a fictional read-only event fixture and make a compact source-to-target dictionary containing the exact field path, raw value, explanation, language tag, reviewer and evidence reference. Have one person who did not write the text trace the same object from the localized screen to the raw event. Reject the draft if that person encounters a second identity, a changed time, an unsupported business assertion or an unexplained edge.

For the final comparison, save a technical diff of the pre-translation and post-translation exports. Classify each changed byte: display-only prose, approved locale formatting, documented alias, or prohibited protocol mutation. Explain in writing why every accepted change belongs to its class. The workshop passes when the target-language reading is clearer and the engineering reviewer can still independently reconstruct the original run, job, dataset and dependencies. Reuse this audit method for a different domain instead of relying on a memorized phrase.

Workshop 10: An independent two-output job

Construct one job that reads customers and orders but writes separate summaries. Use the explicit lineage facet to avoid a fictional many-to-many graph. Evaluate translations of source, target and dependency independently. Begin with a fictional read-only event fixture and make a compact source-to-target dictionary containing the exact field path, raw value, explanation, language tag, reviewer and evidence reference. Have one person who did not write the text trace the same object from the localized screen to the raw event. Reject the draft if that person encounters a second identity, a changed time, an unsupported business assertion or an unexplained edge.

For the final comparison, save a technical diff of the pre-translation and post-translation exports. Classify each changed byte: display-only prose, approved locale formatting, documented alias, or prohibited protocol mutation. Explain in writing why every accepted change belongs to its class. The workshop passes when the target-language reading is clearer and the engineering reviewer can still independently reconstruct the original run, job, dataset and dependencies. Reuse this audit method for a different domain instead of relying on a memorized phrase.

Workshop 11: A field-level privacy review

Find a column used only for filtering. Describe indirect influence without claiming that the field’s raw value is copied into the final dataset. Do not overclaim masking based on a human label; inspect the recorded transformation. Begin with a fictional read-only event fixture and make a compact source-to-target dictionary containing the exact field path, raw value, explanation, language tag, reviewer and evidence reference. Have one person who did not write the text trace the same object from the localized screen to the raw event. Reject the draft if that person encounters a second identity, a changed time, an unsupported business assertion or an unexplained edge.

For the final comparison, save a technical diff of the pre-translation and post-translation exports. Classify each changed byte: display-only prose, approved locale formatting, documented alias, or prohibited protocol mutation. Explain in writing why every accepted change belongs to its class. The workshop passes when the target-language reading is clearer and the engineering reviewer can still independently reconstruct the original run, job, dataset and dependencies. Reuse this audit method for a different domain instead of relying on a memorized phrase.

Workshop 12: A cross-platform data catalog

Use the same dataset alias in two catalog user interfaces. Test that neither system creates a new lineage node and that differing local display titles still point to the same namespace/name identity. Begin with a fictional read-only event fixture and make a compact source-to-target dictionary containing the exact field path, raw value, explanation, language tag, reviewer and evidence reference. Have one person who did not write the text trace the same object from the localized screen to the raw event. Reject the draft if that person encounters a second identity, a changed time, an unsupported business assertion or an unexplained edge.

For the final comparison, save a technical diff of the pre-translation and post-translation exports. Classify each changed byte: display-only prose, approved locale formatting, documented alias, or prohibited protocol mutation. Explain in writing why every accepted change belongs to its class. The workshop passes when the target-language reading is clearer and the engineering reviewer can still independently reconstruct the original run, job, dataset and dependencies. Reuse this audit method for a different domain instead of relying on a memorized phrase.

Workshop 13: A versioned data snapshot

Translate a report describing snapshot 90214 and a separate report describing a later snapshot. Keep original version tokens and distinguish which run consumed which version rather than calling both “current”. Begin with a fictional read-only event fixture and make a compact source-to-target dictionary containing the exact field path, raw value, explanation, language tag, reviewer and evidence reference. Have one person who did not write the text trace the same object from the localized screen to the raw event. Reject the draft if that person encounters a second identity, a changed time, an unsupported business assertion or an unexplained edge.

For the final comparison, save a technical diff of the pre-translation and post-translation exports. Classify each changed byte: display-only prose, approved locale formatting, documented alias, or prohibited protocol mutation. Explain in writing why every accepted change belongs to its class. The workshop passes when the target-language reading is clearer and the engineering reviewer can still independently reconstruct the original run, job, dataset and dependencies. Reuse this audit method for a different domain instead of relying on a memorized phrase.

Workshop 14: A custom vendor connector

Take a vendor-specific facet with a project prefix and a schema URL. Translate only documented prose fields. Record unresolved semantics where the private schema does not support an interpretation, and block unsupported claims. Begin with a fictional read-only event fixture and make a compact source-to-target dictionary containing the exact field path, raw value, explanation, language tag, reviewer and evidence reference. Have one person who did not write the text trace the same object from the localized screen to the raw event. Reject the draft if that person encounters a second identity, a changed time, an unsupported business assertion or an unexplained edge.

For the final comparison, save a technical diff of the pre-translation and post-translation exports. Classify each changed byte: display-only prose, approved locale formatting, documented alias, or prohibited protocol mutation. Explain in writing why every accepted change belongs to its class. The workshop passes when the target-language reading is clearer and the engineering reviewer can still independently reconstruct the original run, job, dataset and dependencies. Reuse this audit method for a different domain instead of relying on a memorized phrase.

Workshop 15: A source-code provenance panel

Write a localized caption explaining what the repository URI and commit reference show. Preserve the exact URL and commit, and test that the button opens the same revision regardless of selected language. Begin with a fictional read-only event fixture and make a compact source-to-target dictionary containing the exact field path, raw value, explanation, language tag, reviewer and evidence reference. Have one person who did not write the text trace the same object from the localized screen to the raw event. Reject the draft if that person encounters a second identity, a changed time, an unsupported business assertion or an unexplained edge.

For the final comparison, save a technical diff of the pre-translation and post-translation exports. Classify each changed byte: display-only prose, approved locale formatting, documented alias, or prohibited protocol mutation. Explain in writing why every accepted change belongs to its class. The workshop passes when the target-language reading is clearer and the engineering reviewer can still independently reconstruct the original run, job, dataset and dependencies. Reuse this audit method for a different domain instead of relying on a memorized phrase.

Workshop 16: An audit handover

Give a second reviewer a random selection of event fixtures, glossary terms and translation approvals. They must be able to reconstruct the source meaning, identify any fallbacks and pass the graph-equivalence checks without relying on verbal memory. Begin with a fictional read-only event fixture and make a compact source-to-target dictionary containing the exact field path, raw value, explanation, language tag, reviewer and evidence reference. Have one person who did not write the text trace the same object from the localized screen to the raw event. Reject the draft if that person encounters a second identity, a changed time, an unsupported business assertion or an unexplained edge.

For the final comparison, save a technical diff of the pre-translation and post-translation exports. Classify each changed byte: display-only prose, approved locale formatting, documented alias, or prohibited protocol mutation. Explain in writing why every accepted change belongs to its class. The workshop passes when the target-language reading is clearer and the engineering reviewer can still independently reconstruct the original run, job, dataset and dependencies. Reuse this audit method for a different domain instead of relying on a memorized phrase.

Workshop 17: A localization rollback

Inject an intentionally ambiguous translation of an output dataset. Demonstrate that the translated dictionary can roll back independently while all published OpenLineage evidence, metadata events and lineage IDs remain unmodified. Begin with a fictional read-only event fixture and make a compact source-to-target dictionary containing the exact field path, raw value, explanation, language tag, reviewer and evidence reference. Have one person who did not write the text trace the same object from the localized screen to the raw event. Reject the draft if that person encounters a second identity, a changed time, an unsupported business assertion or an unexplained edge.

For the final comparison, save a technical diff of the pre-translation and post-translation exports. Classify each changed byte: display-only prose, approved locale formatting, documented alias, or prohibited protocol mutation. Explain in writing why every accepted change belongs to its class. The workshop passes when the target-language reading is clearer and the engineering reviewer can still independently reconstruct the original run, job, dataset and dependencies. Reuse this audit method for a different domain instead of relying on a memorized phrase.

Workshop 18: An accessible graph explanation

Provide keyboard-reachable lineage nodes, an ordered text description of edges and distinguishable source identifiers. Check that the translated graph alt text states direction and uncertainty accurately without inventing absent links. Begin with a fictional read-only event fixture and make a compact source-to-target dictionary containing the exact field path, raw value, explanation, language tag, reviewer and evidence reference. Have one person who did not write the text trace the same object from the localized screen to the raw event. Reject the draft if that person encounters a second identity, a changed time, an unsupported business assertion or an unexplained edge.

For the final comparison, save a technical diff of the pre-translation and post-translation exports. Classify each changed byte: display-only prose, approved locale formatting, documented alias, or prohibited protocol mutation. Explain in writing why every accepted change belongs to its class. The workshop passes when the target-language reading is clearer and the engineering reviewer can still independently reconstruct the original run, job, dataset and dependencies. Reuse this audit method for a different domain instead of relying on a memorized phrase.

Thirty questions data teams ask about OpenLineage localization

FAQ 1: Can I translate an OpenLineage job name?

Translate its label for readers, but keep the technical namespace and name identical unless a separately authorized engineering migration changes the identity. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 2: What does an OpenLineage RunEvent describe?

It reports information related to a particular execution of a Job, including a run state and associated metadata; it is not merely a static catalog description. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 3: Are eventType values localizable?

Their on-screen labels may be translated. Their protocol values START, RUNNING, COMPLETE, ABORT, FAIL and OTHER should stay unchanged in machine-readable records. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 4: Can I translate the producer URI?

No. Translate the explanation of what the producer is; retain the URI as an evidence reference. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 5: Why is a Job namespace important?

A job name is scoped by its namespace. Removing the namespace can join unrelated jobs or split one job into apparent duplicates. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 6: What is a dataset identity?

The combination of dataset namespace and name identifies a data resource within the model; a friendly title is not itself that identity. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 7: Is an input Dataset the same as an output Dataset?

No. Inputs are consumed and outputs are produced by the represented activity. Reversing them reverses what the lineage graph claims. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 8: Why not translate the JSON wholesale?

It contains enumerations, URIs, field paths, type declarations, timestamps and other structured data. Only selected human-facing prose should enter translation. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 9: What is a facet?

A facet is a named piece of additional metadata attached to a Run, Job or Dataset according to a schema. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 10: Do all facets contain translatable text?

No. Many contain controlled numbers, IDs and code values. Documentation and descriptive fields require context-sensitive decisions. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 11: What is column lineage?

It records which source fields contribute to, or influence, particular output fields. It provides finer granularity than dataset-to-dataset edges. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 12: What is DIRECT transformation?

In column lineage it indicates that an output value is derived from a source field’s value; the precise subtype supplies further detail. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 13: What is INDIRECT transformation?

It indicates a source field influences output selection or arrangement without directly supplying the displayed value, as in filtering. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 14: Can a filter be part of lineage?

Yes. A field used in a filter may have indirect influence. Omitting this from a translation can hide why records appear or disappear. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 15: What is a nominal timestamp?

It can describe the expected time or scheduled window for a run, which may differ from actual event timestamps. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 16: What should I do with timezone conversions?

Keep the canonical timestamp and add a clearly labeled display conversion with its offset or zone, validated around daylight-saving boundaries. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 17: Can I call a passed assertion proof of data quality?

Only within the assertion’s defined scope. It is not proof that all fields and records meet every possible quality requirement. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 18: Is nullCount an error count?

Not necessarily. It counts null values in the measured scope; whether they violate business rules depends on the field and assertion. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 19: What is the difference between producer and owner?

Producer identifies the metadata-emitting software or party, whereas ownership metadata can identify responsibility for a job or data asset. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 20: Why save glossary context?

The same words, such as run, source, view and field, mean different things in SQL, orchestration, business reporting and code provenance. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 21: Can a translated label replace the field name?

It can appear next to the field in the user interface; it should not be substituted for the technical schema key. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 22: What if a translation is unavailable?

Show a clearly marked source-language fallback or coverage status rather than pretending the source text is localized. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 23: How do I check graph equivalence?

Compare the exact source/target entity identities and dependency edges before and after locale selection. Presentation ordering and labels can vary without changing edges. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 24: Does all RunEvent input imply all output dependency?

Not necessarily at a fine grain. Explicit lineage facets and field-level evidence may distinguish which sources feed which targets. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 25: Can I translate SQL embedded in metadata?

Translate a human explanation of the SQL; preserve the query as code and provenance evidence. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 26: What is a custom facet?

A schema-defined extension, normally with a project-specific prefix, whose meaning must be checked with its producer documentation. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 27: What if the source documentation contradicts observed lineage?

Record the discrepancy, assign a domain reviewer and avoid translating the unverified claim as fact until the owner resolves it. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 28: Should a translator have production write access?

Usually not. Favor read-only fixtures, scoped permissions, controlled dictionaries and separate engineering approval for any production changes. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 29: How do I test a multilingual release?

Use snapshots of successful and failed runs, identical technical IDs across locales, exact graph-edge comparison, regression tests for search and exports, and reviewer sign-off. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

FAQ 30: Where should I learn general translation and English skills?

Use the root translation system for cross-domain methodology, the Vocabulary Learning Hub for lexical learning and How English Works for appropriate English-language reasoning; this article remains the OpenLineage technical child. The safest next step is to trace the statement to its underlying field, confirm who defines the field, and test whether the target-language reader reaches the same factual conclusion as a source-language reader.

Practical glossary for multilingual data-lineage readers

OpenLineage. An open specification for describing jobs, runs, datasets and their lineage metadata. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

RunEvent. Event reporting a particular job run and its state or additional metadata. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

JobEvent. Event representing job metadata without assuming a completed execution. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

DatasetEvent. Event representing dataset metadata, which is not equivalent to a job run. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Run. A particular execution of a job with its own run identifier. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Job. An identified process that may consume and produce datasets. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Dataset. An identified data resource such as a table, file, view or topic. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Namespace. The scope that pairs with a name to identify an entity. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

RunId. Identifier linking metadata about a particular execution. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Producer. Metadata emitter identified through a URI. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

SchemaURL. Reference to the machine-readable schema used to interpret the event. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Facet. Named, schema-defined additional information attached to a model entity. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Job facet. Extra metadata associated with a job, such as documentation or lineage. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Run facet. Extra information about execution, scheduling or hierarchy. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Dataset facet. Extra metadata about a dataset, such as schema, version or lineage. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Input facet. Metadata scoped to consumption of a dataset. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Output facet. Metadata scoped to production of a dataset. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

EventType. Controlled run-state token in RunEvent. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

EventTime. Timestamp associated with the emitted event. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Nominal time. Expected or scheduled run window, not always its actual execution time. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Terminal event. An event such as COMPLETE, FAIL or ABORT that closes a run lifecycle. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Input. Resource consumed by the process represented in lineage. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Output. Resource produced by the process represented in lineage. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Schema. Formal description of fields and their types or structure. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Field. Named element in a dataset’s schema. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Field path. Location of a field within nested structure. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Documentation facet. Facet containing human documentation for a job. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

SQL facet. Facet that may include source SQL used by a job. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Column lineage. Field-level dependencies between source and output data. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Direct dependency. A relation in which the output value derives from an input value. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Indirect dependency. Influence on output selection or structure without direct value transfer. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Transformation subtype. More specific category such as identity, aggregation, filter, join or sort. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Data quality metric. Recorded measurement such as row count, null count or byte size. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Data quality assertion. Result of evaluating a specified data-quality condition. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

CaptureDate. Time at which a particular recorded metric was captured when provided. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Dataset version. Reference to a particular state of a data resource. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Lifecycle event. Information about a dataset being created, renamed, dropped or otherwise changed. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Lineage graph. Directed representation of data-resource and process dependencies. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Explicit lineage facet. Detailed edges that distinguish genuine dependencies from naive pairwise assumptions. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Alias. Human-readable or alternate reference tied to the same technical identity. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Display dictionary. Reviewed mapping from a source field or phrase to locale-specific reader text. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Fallback. Use of source-language content when no approved target translation exists. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Business glossary. Definitions of business concepts that may span several technical fields. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Regression fixture. A saved event and expected outcome used to detect changed behavior. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Graph equivalence. Equality of underlying entities and dependency relationships across displays. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Schema validation. Check that an event complies with the referenced technical schema. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Provenance. Evidence of the origin, production and history of metadata or data. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Reversibility. Ability to undo a localization change without changing source events. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Information leakage. Exposure of protected technical or business details through translation workflow. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Semantic invariant. A fact that must remain unchanged across languages. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Target-language review. Independent check of linguistic accuracy and domain truth. In translation, state the exact system referent before choosing a friendly phrase; a term may be an interface label, a structured token or both, and the handling differs.

Editorial and technical quality-assurance checklist

• Opening paragraphs answer exactly how to localize OpenLineage, explain the hidden failure and provide the repeatable method.
• Machine record remains unchanged; title, excerpt, documentation and permitted interface prose live in separate presentation fields.
• All synthetic fixtures are identified as fictional and are not represented as empirical evidence.
• Every identity-bearing example includes its raw field path and a way to test the original reference.
• Events are not relabeled as different run outcomes; terminal states remain distinct.
• Time-zone displays are explicitly derived from untouched canonical timestamps.
• Graph direction and column-level direct or indirect dependencies are reviewed independently.
• All quality claims specify exact metric or assertion, scope and source evidence.
• Custom facets are not falsely described as normative standard fields.
• Source URLs, schema pointers, SQL and code references remain trustworthy and legible.
• Links to protected educational owners are used only where genuinely helpful, without changing those pages.
• Release materials contain dictionary version, reviewer, test evidence, known limitations and rollback procedure.

Sources, authority and current standards notes

This article was prepared with the official public OpenLineage specification documentation checked on 1 November 2026. The source documentation distinguishes Run, Job and Dataset objects, the event lifecycle, facet placement, schema descriptions and column-level or explicit lineage. A current documentation page shows a 1.55.0 version label; that label refers to the project’s documentation/release line and should not be confused with every individual facet’s JSON schema revision. Implementations must validate against the schemaURL actually present in their events.

• OpenLineage — Object Model
• OpenLineage — Run Cycle
• OpenLineage — Facets and Extensibility
• OpenLineage — Schema Dataset Facet
• OpenLineage — Column Level Lineage
• OpenLineage — Lineage Job Facet
• OpenLineage — Lineage Dataset Facet
• OpenLineage — Data Quality Metrics
• OpenLineage — Job Documentation Facet

Next routes: return to the Master Art of Translation and study adjacent skills deliberately

This specialized chapter belongs to the Master Art of Translation | The Complete System for Moving Meaning Between Languages. Return there for the overall method of moving meaning between languages, deciding what is invariant and what must be rewritten for people, and selecting a new domain without mixing its technical contracts with this one.

For general methods of learning words and maintaining an accurate term register, visit the protected Vocabulary Learning Hub. For questions about the structure, nuance, register and interpretation of English explanations, the existing How English Works is the appropriate owner. These links are learning routes only: this chapter does not change their definitions, navigation or content.

The final principle is simple: a successful multilingual lineage system lets readers use their own language to ask better questions of the same evidence. A translation may change the wording used to explain a run, job, dataset, schema field, quality metric or dependency. It may not silently change which object is identified, which event happened, what a number measured or where the data came from. Preserve those distinctions, and both the language and the lineage remain trustworthy.