Master Art of Translation | The BIDS 1.11.2 Neuroimaging Dataset Translation System — How Filenames, Sidecars, Events, Phenotypes and Provenance Keep Their Meaning

How do you translate a BIDS neuroimaging dataset, its JSON sidecar metadata, participants.tsv data dictionary, events.tsv task annotations and MRI or EEG study documentation without breaking reproducibility? A safe BIDS translation workflow separates the words a person reads from the filenames, metadata keys, identifiers, time units, coded values and links that software interprets. For a dataset described under Brain Imaging Data Structure (BIDS) 1.11.2, success means a researcher using the translated explanation locates the same subjects and sessions, understands the same measurements, and runs the same analysis against the same data.

The hidden problem is that an apparently excellent translation can damage research evidence. Renaming task-rest, translating a JSON field key, flipping a response category, converting seconds to milliseconds without converting numbers, or treating a derivative as a raw acquisition can make a valid dataset misleading. Even when files still parse, a bilingual reader may draw a different inference about a participant, an event, a physiological signal or a study limitation. Metadata translation therefore requires linguistic judgment plus technical validation, provenance checks, privacy review and a release record.

This complete guide to BIDS dataset metadata translation and multilingual neuroimaging documentation explains how to localize descriptions while keeping BIDS 1.11.2 structure intact. Start with a frozen, permission-cleared source dataset and version inventory. Classify each field as executable identifier, controlled value, numeric measurement, source quotation, human-readable explanation or protected personal information. Translate only the authorized human-facing layer; compare sidecars, events, phenotype descriptions and linked files; run schema and cross-reference checks; and ask a domain reviewer to verify that both language versions support the same scientific decisions.

• Core search questions: how to translate BIDS sidecar JSON, events.tsv, participants.tsv, phenotype metadata and neuroimaging dataset documentation safely
• Scope: research-data localization and multilingual metadata, not clinical imaging interpretation or diagnostic translation
• Architecture: EDKSG-TRANS-MASTER-WORLD-930, an additive child after the 920 OpenAPI contract article
• Principle: human descriptions may be localized; executable keys, paths, scientific values and evidentiary provenance must remain consistent
• Primary controls: declared BIDS version, type-aware mapping, source provenance, privacy review, validator, cross-file joins and bilingual evidence audit
• No geographic limitation: applicable to international research collections and distributed teams

Definition Lock: What BIDS Neuroimaging Translation Means

BIDS, the Brain Imaging Data Structure, is a community specification for organizing and describing neuroimaging and related experimental data. This article uses the official stable 1.11.2 documentation consulted on 1 November 2026. It is not a claim that every historical archive was produced under that edition; the dataset’s declared BIDSVersion controls what a released record claims to implement. BIDS defines filenames and directories, metadata, tabular files, modality rules and derivatives. The data may include MRI, EEG, MEG, iEEG, PET, behavioral measures and other supported modalities.

Translation in this child means controlled cross-language transfer of scientific interpretation, documented definitions and reader actions while respecting immutable machine identities. It does not mean rewriting the binary imaging arrays, changing a participant’s experimental result, changing a medical diagnosis, overriding ethical review or inventing values a dataset does not contain. A text may be understandable in one language and inaccurate in the study context; the review target is the meaning of the actual recorded experiment.

Search Intent and the Reader’s Real Task

The most useful questions are practical: Can a team translate BIDS metadata? Which JSON values are safe to localize? How should a multilingual research repository display task names? Can events and questionnaire response codes be translated? What does validation prove, and what can it miss? How should data dictionaries distinguish sex, gender, missing responses, sessions and units? When should a translator consult a neuroscientist or a data steward? This guide answers them with tests instead of asserting that a generic translation model understands every field.

Three readers must agree after localization. The dataset engineer must still parse and address the same files. The investigator must still infer the same experiment and quantities. The participant-facing governance reviewer must still see only the information authorized for use and release. Failure for any one reader is a meaningful defect, even if the target prose is fluent.

The Hidden Problem: Language Is Embedded in Evidence

A filename such as sub-01_task-rest_bold.nii.gz looks readable because it contains recognizable words. But it is an identifier formed from entities and a suffix. It is not a paragraph. A JSON object likewise has key names that programs read and values with different type constraints; some values are free text and others are controlled codes, numbers, lists or references. The same distinction appears in TSV tables, where participant_id is a join key and a level label can be a factor in an analysis model.

The next layer is scientific interpretation. For example, BOLD MRI concerns a hemodynamic signal; it is not a direct measure of the number of thoughts. Onset in an event table can be relative to a recording; it is not automatically a wall-clock date. A derivative is a pipeline output, not another acquisition. These boundaries are easier to preserve when the translation team owns a field contract that specifies which characters may change, which values must remain equal, and what evidence resolves ambiguity.

Mechanism: Six Layers of Meaning That Must Travel Together

• Lexical layer: what a researcher understands by task, response, session, event, amplitude and signal
• Schema layer: the literal keys, types, controlled codes, folder names and parser expectations
• Measurement layer: numbers, units, precision, ordering, coordinate frames and time origins
• Relational layer: subject/run joins, sidecar inheritance, stimulus references and derivative source links
• Provenance layer: acquisition context, data-processing history, software versions, citations and release records
• Governance layer: participant privacy, permitted processing, access, withdrawal, license and attribution

A good translation can change the first layer without changing the other five. Sometimes an authorized scientific revision must change the underlying data; that requires an independently governed, versioned correction with clear evidence. It is not a translator’s silent editorial privilege. The important separation is between a localized presentation of a record and an alteration of the record itself.

Repeatable 18-Gate Production Method

1. Define the requested use

State whether the output is a translated study summary, a bilingual data dictionary, a repository interface or participant-facing documentation. Do not presume that all BIDS files should be rewritten. The release gate should record a concrete pass, an unresolved issue with owner, or a deliberate rejection. A reassuring label is not evidence. Save a trace to the exact source version and the item reviewed so another researcher can reproduce the decision.

2. Freeze and hash the source

Record repository version, release date, checksums, licenses and access scope before copying or processing information. The release gate should record a concrete pass, an unresolved issue with owner, or a deliberate rejection. A reassuring label is not evidence. Save a trace to the exact source version and the item reviewed so another researcher can reproduce the decision.

3. Record the BIDS edition

Read the exact BIDSVersion claim and map rules to that release; do not automatically upgrade metadata to the latest published specification. The release gate should record a concrete pass, an unresolved issue with owner, or a deliberate rejection. A reassuring label is not evidence. Save a trace to the exact source version and the item reviewed so another researcher can reproduce the decision.

4. Inventory every file class

List README, CHANGES, dataset description, sidecars, participants, events, phenotype, stimuli, sourcedata and derivatives as applicable. The release gate should record a concrete pass, an unresolved issue with owner, or a deliberate rejection. A reassuring label is not evidence. Save a trace to the exact source version and the item reviewed so another researcher can reproduce the decision.

5. Classify translatable surfaces

Mark fields as machine key, controlled value, free text, measurement, identifier, link, source quotation or sensitive personal material. The release gate should record a concrete pass, an unresolved issue with owner, or a deliberate rejection. A reassuring label is not evidence. Save a trace to the exact source version and the item reviewed so another researcher can reproduce the decision.

6. Protect identity

Freeze subject/session/run keys, dataset identifiers, path references, DOI links, software names and exact original names where identity is at stake. The release gate should record a concrete pass, an unresolved issue with owner, or a deliberate rejection. A reassuring label is not evidence. Save a trace to the exact source version and the item reviewed so another researcher can reproduce the decision.

7. Define the termbase

Record semantic distinctions such as session versus visit, event versus trial, observed versus derived, reference versus coordinate frame. The release gate should record a concrete pass, an unresolved issue with owner, or a deliberate rejection. A reassuring label is not evidence. Save a trace to the exact source version and the item reviewed so another researcher can reproduce the decision.

8. Build a bilingual field ledger

For each candidate string retain source location, source text, context, proposed translation, decision, reviewer and acceptance test. The release gate should record a concrete pass, an unresolved issue with owner, or a deliberate rejection. A reassuring label is not evidence. Save a trace to the exact source version and the item reviewed so another researcher can reproduce the decision.

9. Obtain domain clarification

Escalate underspecified task conditions, instrument meaning, code levels, coordinate axes, missing-value semantics and statistical claims. The release gate should record a concrete pass, an unresolved issue with owner, or a deliberate rejection. A reassuring label is not evidence. Save a trace to the exact source version and the item reviewed so another researcher can reproduce the decision.

10. Translate safely

Write natural target-language explanations without changing keys, numbers, unit conventions, controlled vocabularies or hidden study assumptions. The release gate should record a concrete pass, an unresolved issue with owner, or a deliberate rejection. A reassuring label is not evidence. Save a trace to the exact source version and the item reviewed so another researcher can reproduce the decision.

11. Validate syntax

Parse JSON and TSV and check filenames, types, schemas, reserved values and supported modality conventions. The release gate should record a concrete pass, an unresolved issue with owner, or a deliberate rejection. A reassuring label is not evidence. Save a trace to the exact source version and the item reviewed so another researcher can reproduce the decision.

12. Validate relationships

Resolve all participant joins, stimulus links, fieldmap targets, inheritance and derivative source references. The release gate should record a concrete pass, an unresolved issue with owner, or a deliberate rejection. A reassuring label is not evidence. Save a trace to the exact source version and the item reviewed so another researcher can reproduce the decision.

13. Verify meaning

Conduct bilingual, field-level comparison of task instructions, study design, measurement definitions and uncertainty language. The release gate should record a concrete pass, an unresolved issue with owner, or a deliberate rejection. A reassuring label is not evidence. Save a trace to the exact source version and the item reviewed so another researcher can reproduce the decision.

14. Verify privacy

Inspect personal information, quotations, free text, raw source exports, tool access and rights for public display. The release gate should record a concrete pass, an unresolved issue with owner, or a deliberate rejection. A reassuring label is not evidence. Save a trace to the exact source version and the item reviewed so another researcher can reproduce the decision.

15. Test with a reader

Ask an unfamiliar qualified reader to perform realistic actions using only the translated documentation. The release gate should record a concrete pass, an unresolved issue with owner, or a deliberate rejection. A reassuring label is not evidence. Save a trace to the exact source version and the item reviewed so another researcher can reproduce the decision.

16. Inspect accessibility

Check language tags, table headings, screen-reader order, plot captions, bidirectional identifiers and text expansion. The release gate should record a concrete pass, an unresolved issue with owner, or a deliberate rejection. A reassuring label is not evidence. Save a trace to the exact source version and the item reviewed so another researcher can reproduce the decision.

17. Publish with provenance

Release a stable language-specific display layer with reference to the canonical dataset, reviewed date and issue-contact route. The release gate should record a concrete pass, an unresolved issue with owner, or a deliberate rejection. A reassuring label is not evidence. Save a trace to the exact source version and the item reviewed so another researcher can reproduce the decision.

18. Recheck after changes

Tie each source update to the target variants; audit stale translations, changed code meanings and downstream cached exports. The release gate should record a concrete pass, an unresolved issue with owner, or a deliberate rejection. A reassuring label is not evidence. Save a trace to the exact source version and the item reviewed so another researcher can reproduce the decision.

Failure Modes: What a Fluent Translation Can Still Get Wrong

Renaming a key

A JSON parser looks for literal names. Localize the description, not the key. The diagnostic should identify the original field, the translator’s proposed alteration, the actual consumer of that field, and the consequence. Fix the mismatch at the proper layer: presentation, metadata correction, processing pipeline or governance decision. Record who authorized the repair and whether any downstream publication must be amended.

Rewriting a file path

A cross-reference is an address. Translate a label for the address, not its stored bytes. The diagnostic should identify the original field, the translator’s proposed alteration, the actual consumer of that field, and the consequence. Fix the mismatch at the proper layer: presentation, metadata correction, processing pipeline or governance decision. Record who authorized the repair and whether any downstream publication must be amended.

Flattening missingness

Not observed, not applicable and a measured zero are different statistical states. The diagnostic should identify the original field, the translator’s proposed alteration, the actual consumer of that field, and the consequence. Fix the mismatch at the proper layer: presentation, metadata correction, processing pipeline or governance decision. Record who authorized the repair and whether any downstream publication must be amended.

Ignoring the time origin

Seconds after acquisition start do not become a local timestamp by translation. The diagnostic should identify the original field, the translator’s proposed alteration, the actual consumer of that field, and the consequence. Fix the mismatch at the proper layer: presentation, metadata correction, processing pipeline or governance decision. Record who authorized the repair and whether any downstream publication must be amended.

Swapping experimental factors

A trial label used in a design matrix cannot be casually reclassified for readability. The diagnostic should identify the original field, the translator’s proposed alteration, the actual consumer of that field, and the consequence. Fix the mismatch at the proper layer: presentation, metadata correction, processing pipeline or governance decision. Record who authorized the repair and whether any downstream publication must be amended.

Changing an atlas index

A region name must remain paired with the same integer and spatial reference. The diagnostic should identify the original field, the translator’s proposed alteration, the actual consumer of that field, and the consequence. Fix the mismatch at the proper layer: presentation, metadata correction, processing pipeline or governance decision. Record who authorized the repair and whether any downstream publication must be amended.

Assuming consent from availability

Repository access does not itself prove that data may be submitted to external translation services. The diagnostic should identify the original field, the translator’s proposed alteration, the actual consumer of that field, and the consequence. Fix the mismatch at the proper layer: presentation, metadata correction, processing pipeline or governance decision. Record who authorized the repair and whether any downstream publication must be amended.

Conflating raw and processed data

A filtered image is not the same evidentiary layer as the input acquisition. The diagnostic should identify the original field, the translator’s proposed alteration, the actual consumer of that field, and the consequence. Fix the mismatch at the proper layer: presentation, metadata correction, processing pipeline or governance decision. Record who authorized the repair and whether any downstream publication must be amended.

Overpromising a validator

Machine syntax and metadata conformance checks do not establish scientific or linguistic truth. The diagnostic should identify the original field, the translator’s proposed alteration, the actual consumer of that field, and the consequence. Fix the mismatch at the proper layer: presentation, metadata correction, processing pipeline or governance decision. Record who authorized the repair and whether any downstream publication must be amended.

Treating quality warnings as diagnoses

A noisy electrode or motion-affected image is a measurement-quality matter, not a clinical statement. The diagnostic should identify the original field, the translator’s proposed alteration, the actual consumer of that field, and the consequence. Fix the mismatch at the proper layer: presentation, metadata correction, processing pipeline or governance decision. Record who authorized the repair and whether any downstream publication must be amended.

Erasing source-language quotation

Research interpretation can depend on what participants actually saw or said. The diagnostic should identify the original field, the translator’s proposed alteration, the actual consumer of that field, and the consequence. Fix the mismatch at the proper layer: presentation, metadata correction, processing pipeline or governance decision. Record who authorized the repair and whether any downstream publication must be amended.

Declaring a new edition without migration

A new published specification does not retroactively modify old recorded data. The diagnostic should identify the original field, the translator’s proposed alteration, the actual consumer of that field, and the consequence. Fix the mismatch at the proper layer: presentation, metadata correction, processing pipeline or governance decision. Record who authorized the repair and whether any downstream publication must be amended.

Field-Level Evidence Workshop 1: Dataset identity, descriptions, standards and version boundaries

A BIDS dataset is a structured research record, not a collection of prose files. Establish which exact version of the specification the deposited dataset claims to follow before deciding what a string may mean. The same screen may show ordinary descriptions next to controlled keys, repository identifiers, licensing statements and strings copied from the analysis environment. Translation review should be keyed to the actual release, not a vague statement that the data are BIDS-compatible.

Use this module as a field dossier, not as a universal substitute for versioned specification text. In each fictional example, the source laboratory has the duty to define what was measured, the data steward controls authorized changes, the translator handles cross-language expression, and an independent analyst checks whether the same decision can be made from the target version. The examples are intentionally illustrative; they are not real participant records, release manifests or clinical advice.

Diagnostic 001: dataset_description.json Name

Source situation. The Name value describes the study for people; the JSON key remains Name and the published study title may have a citation identity. In this fictional BIDS dataset, the translator encounters dataset_description.json Name while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. An editor translates the field key or silently replaces the cited project title with a marketing-style alternative. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Retain the exact key, record the canonical title, and add a localized display title only in an explicitly documented layer. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The dataset description, DOI landing page and original contributor-approved title establish identity. A practical acceptance test is: Compare machine parsing, citation text and visible multilingual record. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Inspect the object as both a scientist and a file parser. The human account must not rely on an assumption that the schema cannot support. If the two descriptions disagree, pause the release and assign the issue to a domain owner.

Diagnostic 002: BIDSVersion value

Source situation. BIDSVersion records the specification version declared by the dataset, which may differ from the newest public specification. In this fictional BIDS dataset, the translator encounters BIDSVersion value while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A localization sweep replaces an older version number with 1.11.2 merely because that edition is current. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Preserve the actual declared value; upgrade only through a separate documented technical conformance process. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The original versioned record and validator report identify the implemented rules. A practical acceptance test is: Diff the value against the archived source and rerun a validator for the declared version. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Run a reversible test rather than editing the only available copy. Preserve the canonical source, produce an isolated target-language view, and compare what each version causes a reader or program to select.

Diagnostic 003: DatasetType raw or derivative

Source situation. DatasetType categorizes a dataset as raw or derivative where applicable, affecting interpretation of processing provenance. In this fictional BIDS dataset, the translator encounters DatasetType raw or derivative while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A prose translation changes raw into unprocessed in a field that requires a controlled machine value. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep the machine value and translate explanatory help beside it with a note about what conversion steps may still have occurred. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The dataset tree, acquisition files and pipeline records show whether the package is raw or derived. A practical acceptance test is: Check schema compatibility and the downstream pipeline’s dispatch behavior. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Separate observation from interpretation. A code or measured number belongs to the original record; a sentence explaining that value may be revised, but the revision must be traceable and scientifically warranted.

Diagnostic 004: GeneratedBy metadata

Source situation. A derivative may identify generating software, version, container and associated processing provenance. In this fictional BIDS dataset, the translator encounters GeneratedBy metadata while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translator normalizes a tool name, translates a software version or drops a container reference as irrelevant jargon. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Freeze executable names and versions; explain what the tool did in natural language without changing its reported identity. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. Pipeline log, container digest and release notes provide traceable provenance. A practical acceptance test is: Compare a sample derivative report with the actual tool invocation and recorded release. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Include the negative case in review. Ask what a researcher could falsely infer when the translation omits a qualification, and whether that inference would affect a published conclusion or participant treatment.

Diagnostic 005: SourceDatasets links

Source situation. SourceDatasets identifies upstream datasets used to create a derivative and may include URLs or other references. In this fictional BIDS dataset, the translator encounters SourceDatasets links while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. The target-language article replaces a source identifier with a translated human name that cannot be resolved. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep source dataset identifiers intact and translate only the associated explanatory labels. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The archived input manifest and upstream repository record resolve the dependency. A practical acceptance test is: Follow each reference and confirm it resolves to the same data version. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Use an independent reviewer who can explain the experimental process rather than merely certify fluency. Give that reviewer the source field, its neighboring fields and a small reproducible sample.

Diagnostic 006: License and consent scope

Source situation. A dataset license and human-participant consent are separate questions even when a description calls the material open. In this fictional BIDS dataset, the translator encounters License and consent scope while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. An adaptation makes public download sound like unrestricted consent to redistribute identifiable recordings. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Quote the actual license reference, retain usage conditions and request governance review for sensitive data. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The repository access conditions, institutional review approvals and consent policy define permitted use. A practical acceptance test is: Ask a reviewer whether the translation grants an action the source does not permit. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Treat semantic equivalence as an evidence question. Identify the experiment, provenance or versioned specification that would disprove a tempting but incorrect paraphrase.

Diagnostic 007: Dataset author and funding fields

Source situation. Authors, acknowledgements, funding bodies and grants have attribution and persistent-identity functions. In this fictional BIDS dataset, the translator encounters Dataset author and funding fields while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. Editors localize organization names freely or expand an acronym without linking it to the responsible funder. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Preserve the recorded author attribution and funding identifier; translate explanatory prose without changing credit. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The deposited citation, contributor identifiers and award records are authoritative. A practical acceptance test is: Reconcile the translated record against the same citation and funder index. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Record the test at the correct level. A JSON parse is useful for keys and types; it cannot determine whether a cognitive condition has been translated into the wrong construct.

Diagnostic 008: Study README and CHANGES

Source situation. README explains interpretation and CHANGES describes actual revisions across releases. In this fictional BIDS dataset, the translator encounters Study README and CHANGES while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. An editor converts a source correction note into a claim that the underlying study measurement itself was corrected. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Preserve timestamps, versions and scope of each change and explain whether only documentation changed. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. A commit diff and repository release record show precisely which files moved. A practical acceptance test is: Have a second reader distinguish revised description from revised participant observations. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Check portability and accessibility together. A target-language reader should be able to copy a code, locate a file and understand an uncertainty note without relying on hidden English instructions.

Transfer Exercise 1: From Field Repair to Release Decision

Take two cases in dataset identity, descriptions, standards and version boundaries and build a miniature bilingual review bundle. Include the literal field path, a valid fictional source record, the original explanation, a candidate target explanation, a deliberately faulty alternative and the program or human decision affected. Then create a one-page disposition table: accept the translation, revise the display text, require a dataset correction, request a scientist’s answer, or withhold the material because of privacy. The important skill is knowing where to act. Localization belongs in the display layer when identity and values are fixed; correcting recorded evidence belongs in a governed release with transparent provenance.

Field-Level Evidence Workshop 2: Participants, sessions, privacy and consent-preserving descriptions

Human-facing translation is particularly dangerous when the field describes people rather than instruments. A participant code is not a name. A session may group related acquisitions without being a one-to-one calendar visit. Consent and privacy are decisions made under an actual protocol, not features inferred from the fact that a string is translatable. Use the authorized source record and minimal exposure.

Use this module as a field dossier, not as a universal substitute for versioned specification text. In each fictional example, the source laboratory has the duty to define what was measured, the data steward controls authorized changes, the translator handles cross-language expression, and an independent analyst checks whether the same decision can be made from the target version. The examples are intentionally illustrative; they are not real participant records, release manifests or clinical advice.

Diagnostic 009: sub- participant labels

Source situation. The sub- label is an analysis identifier, often a pseudonym assigned without public meaning. In this fictional BIDS dataset, the translator encounters sub- participant labels while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translator expands sub-011 into a real first name or translates the subject label into a descriptive category. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Do not relabel participants or infer identity; keep stable pseudonyms and place approved descriptions elsewhere. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The subject key mapping held by authorized data stewards determines identity and is not publication material. A practical acceptance test is: Cross-check all filename references after localization without consulting private identifiers unnecessarily. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Run a reversible test rather than editing the only available copy. Preserve the canonical source, produce an isolated target-language view, and compare what each version causes a reader or program to select.

Diagnostic 010: sessions.tsv session grouping

Source situation. A session records a logical grouping of acquisition; it may not correspond to one hospital appointment. In this fictional BIDS dataset, the translator encounters sessions.tsv session grouping while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translated session label equates ses-02 with a second visit despite merged acquisition days. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep ses- identifiers; describe the actual collection design from the research protocol. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The acquisition schedule, methods file and scans.tsv date pattern show grouping. A practical acceptance test is: Ensure readers can reconstruct the same session/run mapping. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Separate observation from interpretation. A code or measured number belongs to the original record; a sentence explaining that value may be revised, but the revision must be traceable and scientifically warranted.

Diagnostic 011: participants.tsv age

Source situation. Age may be measured in years, rounded, shifted or withheld for privacy, depending on documented procedures. In this fictional BIDS dataset, the translator encounters participants.tsv age while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. The target description says exact age at scan even when values are rounded or top-coded. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Document precision, reference time and confidentiality treatment without reconstructing identities. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The study data dictionary and de-identification plan specify age derivation. A practical acceptance test is: Check that translated labels do not imply a more precise measurement than released. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Include the negative case in review. Ask what a researcher could falsely infer when the translation omits a qualification, and whether that inference would affect a published conclusion or participant treatment.

Diagnostic 012: participant sex and gender fields

Source situation. A recorded sex field or gender questionnaire response has the operational meaning specified by its study instrument. In this fictional BIDS dataset, the translator encounters participant sex and gender fields while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translator substitutes one concept for the other or rewrites category labels to satisfy a generic glossary. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Preserve the construct actually measured, retain coded values and review inclusive wording with study governance. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The instrument wording, variable documentation and participant-facing materials establish the construct. A practical acceptance test is: Test whether the translation would lead analysts to group different observations. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Use an independent reviewer who can explain the experimental process rather than merely certify fluency. Give that reviewer the source field, its neighboring fields and a small reproducible sample.

Diagnostic 013: control and patient group codes

Source situation. A study group identifier is a classification assigned under protocol-defined inclusion criteria. In this fictional BIDS dataset, the translator encounters control and patient group codes while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A reader-facing translation turns a research group code into a diagnosis for every participant. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep codes immutable and render an approved description of how membership was established. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The recruitment protocol and group-level inclusion rules are the evidence, not assumptions from labels. A practical acceptance test is: Verify that the target wording does not infer an individual’s clinical status. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Treat semantic equivalence as an evidence question. Identify the experiment, provenance or versioned specification that would disprove a tempting but incorrect paraphrase.

Diagnostic 014: phenotype sensitive responses

Source situation. Questionnaire and phenotype rows may reveal cognitive, health or behavioral information even after naming is removed. In this fictional BIDS dataset, the translator encounters phenotype sensitive responses while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translator sends identifiable free text into an unauthorized cloud translation service. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep processing inside approved environments, minimize exposure and use data use agreements where applicable. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The ethics approval and data management plan specify allowed processors and retention. A practical acceptance test is: Audit who accessed records, which service received them and whether copies persist. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Record the test at the correct level. A JSON parse is useful for keys and types; it cannot determine whether a cognitive condition has been translated into the wrong construct.

Diagnostic 015: multi-site participant labels

Source situation. Two contributing studies may each use sub-001 while retaining distinct local participant keys. In this fictional BIDS dataset, the translator encounters multi-site participant labels while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. Consolidation in another language silently merges subjects based on matching labels. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Maintain dataset namespaces and document any authorized harmonized mapping separately. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. Site-specific acquisition logs and repository manifests show separate subject identities. A practical acceptance test is: Run duplicate checks using the intended composite identifiers. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Check portability and accessibility together. A target-language reader should be able to copy a code, locate a file and understand an uncertainty note without relying on hidden English instructions.

Diagnostic 016: withdrawn or restricted records

Source situation. Some records must be excluded or restricted after a participant revokes permission under applicable rules. In this fictional BIDS dataset, the translator encounters withdrawn or restricted records while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translated public index continues advertising withdrawn events or scans after the source has been corrected. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Synchronize access rights and translations through a controlled withdrawal-and-cache invalidation procedure. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The repository’s release log and governance order establish the record’s current status. A practical acceptance test is: Inspect search results, downloadable snapshots and derivative catalogs after withdrawal. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Inspect the object as both a scientist and a file parser. The human account must not rely on an assumption that the schema cannot support. If the two descriptions disagree, pause the release and assign the issue to a domain owner.

Transfer Exercise 2: From Field Repair to Release Decision

Take two cases in participants, sessions, privacy and consent-preserving descriptions and build a miniature bilingual review bundle. Include the literal field path, a valid fictional source record, the original explanation, a candidate target explanation, a deliberately faulty alternative and the program or human decision affected. Then create a one-page disposition table: accept the translation, revise the display text, require a dataset correction, request a scientist’s answer, or withhold the material because of privacy. The important skill is knowing where to act. Localization belongs in the display layer when identity and values are fixed; correcting recorded evidence belongs in a governed release with transparent provenance.

Field-Level Evidence Workshop 3: Filenames, directory labels, paths and case-sensitive identifiers

BIDS file naming combines key–value entities with suffixes that scientific software reads. Translators must separate stable machine names from human labels. A filename is not a sentence; a familiar English word inside it may be part of a protocol. Most safe multilingual output keeps the file tree identical and presents a translated view or data dictionary alongside it.

Use this module as a field dossier, not as a universal substitute for versioned specification text. In each fictional example, the source laboratory has the duty to define what was measured, the data steward controls authorized changes, the translator handles cross-language expression, and an independent analyst checks whether the same decision can be made from the target version. The examples are intentionally illustrative; they are not real participant records, release manifests or clinical advice.

Diagnostic 017: task-rest filename

Source situation. The task-rest entity identifies a rest acquisition; the spelling matters to parsers and scripts. In this fictional BIDS dataset, the translator encounters task-rest filename while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translator renames task-rest to task-repos while leaving scripts set to task-rest. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep machine filenames unchanged and explain resting-state acquisition in localized metadata. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The acquisition protocol and BIDS naming grammar specify the intended task. A practical acceptance test is: Run the same selector against the original and localized package. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Separate observation from interpretation. A code or measured number belongs to the original record; a sentence explaining that value may be revised, but the revision must be traceable and scientifically warranted.

Diagnostic 018: run-01 and run-02

Source situation. Run identifiers distinguish separate uninterrupted acquisitions rather than quality ranking. In this fictional BIDS dataset, the translator encounters run-01 and run-02 while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A caption calls run-02 the improved scan and swaps numerical labels to match narration order. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep run indices as acquired and explain restart or order from logs. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The scanner sequence list and behavioral timing logs define runs. A practical acceptance test is: Match each run to the same raw data checksum. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Include the negative case in review. Ask what a researcher could falsely infer when the translation omits a qualification, and whether that inference would affect a published conclusion or participant treatment.

Diagnostic 019: anat versus func directories

Source situation. anat and func are datatype-directory names used for different forms of MRI data. In this fictional BIDS dataset, the translator encounters anat versus func directories while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A designer translates directory names to anatomical or functional words inside the downloaded archive. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep directory tokens stable; localize only explanatory navigation labels. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The BIDS datatype list establishes allowed directory vocabulary. A practical acceptance test is: Open the package with a BIDS-aware tool and check expected files appear. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Use an independent reviewer who can explain the experimental process rather than merely certify fluency. Give that reviewer the source field, its neighboring fields and a small reproducible sample.

Diagnostic 020: task label collision

Source situation. Task labels must remain distinct when case is ignored so different files do not collapse on some filesystems. In this fictional BIDS dataset, the translator encounters task label collision while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translation produces task-Go and task-go for distinct original tasks. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Create approved short stable labels and test filenames on case-insensitive systems. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The naming plan and task specification determine intended distinctions. A practical acceptance test is: Scan for collisions with case-folded path comparisons. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Treat semantic equivalence as an evidence question. Identify the experiment, provenance or versioned specification that would disprove a tempting but incorrect paraphrase.

Diagnostic 021: entity ordering

Source situation. BIDS specifies conventions for the order of filename entities. In this fictional BIDS dataset, the translator encounters entity ordering while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A copyeditor rearranges labels so the filename reads more naturally as a phrase. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Maintain the valid entity order, moving descriptions only into display text. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The version-matched entity table and existing valid files anchor ordering. A practical acceptance test is: Validate representative filenames and compare sample pipeline glob patterns. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Record the test at the correct level. A JSON parse is useful for keys and types; it cannot determine whether a cognitive condition has been translated into the wrong construct.

Diagnostic 022: file suffix bold

Source situation. The bold suffix denotes a type of functional MRI acquisition and is not a general adjective. In this fictional BIDS dataset, the translator encounters file suffix bold while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translator expands or changes the suffix to localized prose within the data package. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Freeze the suffix and provide an accessible modality definition in a glossary. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The modality-specific specification and scan type indicate file kind. A practical acceptance test is: Check the workflow correctly discovers the same NIfTI volumes. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Check portability and accessibility together. A target-language reader should be able to copy a code, locate a file and understand an uncertainty note without relying on hidden English instructions.

Diagnostic 023: extension nii.gz

Source situation. The nii.gz extension describes compressed NIfTI data; the compression is not an optional stylistic flourish. In this fictional BIDS dataset, the translator encounters extension nii.gz while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A content translation tool converts paths into .nii or inserts spaces into the extension. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Preserve byte-identical filenames and keep binary images out of text translation tools. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The file format, checksums and original manifest are definitive. A practical acceptance test is: Recompute hashes and ensure decompression and image reading both succeed. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Inspect the object as both a scientist and a file parser. The human account must not rely on an assumption that the schema cannot support. If the two descriptions disagree, pause the release and assign the issue to a domain owner.

Diagnostic 024: BIDS URIs and relative paths

Source situation. A relative path or BIDS URI resolves within dataset-defined boundaries and is not merely a hyperlink label. In this fictional BIDS dataset, the translator encounters BIDS URIs and relative paths while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translator replaces a file path with a friendly translated caption inside a machine reference. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Retain the URI or path as-is and put human-friendly text in a separate field. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The reference target and dataset-relative resolver define link identity. A practical acceptance test is: Resolve each link from the exported multilingual package to the same file. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Run a reversible test rather than editing the only available copy. Preserve the canonical source, produce an isolated target-language view, and compare what each version causes a reader or program to select.

Transfer Exercise 3: From Field Repair to Release Decision

Take two cases in filenames, directory labels, paths and case-sensitive identifiers and build a miniature bilingual review bundle. Include the literal field path, a valid fictional source record, the original explanation, a candidate target explanation, a deliberately faulty alternative and the program or human decision affected. Then create a one-page disposition table: accept the translation, revise the display text, require a dataset correction, request a scientist’s answer, or withhold the material because of privacy. The important skill is knowing where to act. Localization belongs in the display layer when identity and values are fixed; correcting recorded evidence belongs in a governed release with transparent provenance.

Field-Level Evidence Workshop 4: JSON sidecars, inheritance, units and machine-readable constraints

BIDS JSON sidecars can combine normative field names, quantities, booleans, identifiers and explanatory prose. The inheritance principle means one metadata value may apply to several recordings. A safe transformation identifies the effective metadata at the file level instead of only translating the nearest sidecar. A translated explanation must not silently change acquisition numbers or move a value into the wrong level.

Use this module as a field dossier, not as a universal substitute for versioned specification text. In each fictional example, the source laboratory has the duty to define what was measured, the data steward controls authorized changes, the translator handles cross-language expression, and an independent analyst checks whether the same decision can be made from the target version. The examples are intentionally illustrative; they are not real participant records, release manifests or clinical advice.

Diagnostic 025: TaskName display versus task- entity

Source situation. TaskName may provide a human-readable task name while a filename task- label helps identify the acquisition. In this fictional BIDS dataset, the translator encounters TaskName display versus task- entity while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. The translated TaskName suggests a different cognitive task than the unchanged filename and event protocol. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Use an approved translated display with an explicit mapping to the source task definition. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The instruction script and task registration document establish the experiment. A practical acceptance test is: Ask a bilingual domain reviewer to match sample stimuli and task description. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Include the negative case in review. Ask what a researcher could falsely infer when the translation omits a qualification, and whether that inference would affect a published conclusion or participant treatment.

Diagnostic 026: RepetitionTime quantity

Source situation. RepetitionTime is a numeric acquisition parameter conventionally expressed in seconds in BIDS MRI metadata. In this fictional BIDS dataset, the translator encounters RepetitionTime quantity while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translator changes 0.8 to 800 believing the value should be displayed in milliseconds. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep the numeric source and declared convention; convert only for an explicitly labeled explanatory display. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The source JSON and scanner protocol identify the measured interval. A practical acceptance test is: Compare the source value to displayed units and pipeline expectations. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Use an independent reviewer who can explain the experimental process rather than merely certify fluency. Give that reviewer the source field, its neighboring fields and a small reproducible sample.

Diagnostic 027: EchoTime and unit

Source situation. EchoTime describes an echo interval with domain-specific unit convention. In this fictional BIDS dataset, the translator encounters EchoTime and unit while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A localization tool strips a leading zero or changes the decimal separator inside JSON. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Retain machine-readable numeric syntax and keep any localized decimal formatting outside the file. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The acquisition record and BIDS metadata definition specify meaning. A practical acceptance test is: Parse the JSON and compare numeric equality across releases. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Treat semantic equivalence as an evidence question. Identify the experiment, provenance or versioned specification that would disprove a tempting but incorrect paraphrase.

Diagnostic 028: PhaseEncodingDirection axis

Source situation. PhaseEncodingDirection combines an axis index and sign where appropriate; it is not a geographic compass bearing. In this fictional BIDS dataset, the translator encounters PhaseEncodingDirection axis while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. The value j- is translated as left or west without considering image axes and orientation. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Leave the coded direction untouched; explain the coordinate convention with domain review. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. Image geometry, acquisition parameters and specification definitions support interpretation. A practical acceptance test is: Compare distortion-correction outputs and image-axis metadata. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Record the test at the correct level. A JSON parse is useful for keys and types; it cannot determine whether a cognitive condition has been translated into the wrong construct.

Diagnostic 029: IntendedFor fieldmap targets

Source situation. IntendedFor may point to acquisition files for which a fieldmap is intended. In this fictional BIDS dataset, the translator encounters IntendedFor fieldmap targets while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. The translator changes file names in a list but not the actual target files, silently breaking associations. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Preserve references and confirm each points to a matching acquisition. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The fieldmap sidecar, related imaging series and BIDS resolution rules show target mapping. A practical acceptance test is: Resolve and test all intended target paths with a parser. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Check portability and accessibility together. A target-language reader should be able to copy a code, locate a file and understand an uncertainty note without relying on hidden English instructions.

Diagnostic 030: inherited metadata conflicts

Source situation. A JSON sidecar higher in the directory hierarchy can affect multiple more-specific files. In this fictional BIDS dataset, the translator encounters inherited metadata conflicts while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A reviewer reads only one local file and assumes an absent field is missing from the entire dataset. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Compute effective inherited metadata for each file before diagnosing an omission. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The inheritance-principle documentation and complete directory tree supply context. A practical acceptance test is: Use the validator and a file-level effective-metadata inventory. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Inspect the object as both a scientist and a file parser. The human account must not rely on an assumption that the schema cannot support. If the two descriptions disagree, pause the release and assign the issue to a domain owner.

Diagnostic 031: slice timing arrays

Source situation. SliceTiming is an ordered list linked to slice acquisition and cannot be freely sorted for readability. In this fictional BIDS dataset, the translator encounters slice timing arrays while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A spreadsheet editor sorts the values numerically and thus changes the acquisition-time sequence. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Preserve array order and numeric precision; translate description of timing separately. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. Sequence specification and image layout determine interpretation. A practical acceptance test is: Compare array length and index order to the source. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Run a reversible test rather than editing the only available copy. Preserve the canonical source, produce an isolated target-language view, and compare what each version causes a reader or program to select.

Diagnostic 032: JSON booleans and null values

Source situation. true, false and null are syntactic values with meanings defined by individual fields. In this fictional BIDS dataset, the translator encounters JSON booleans and null values while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translation step changes true to a target-language word and converts null to an empty string. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep JSON literals and document the difference between missing, unknown and false where the field defines it. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The JSON parser and BIDS metadata schema define valid value types. A practical acceptance test is: Run type checks and assess three-way missingness in the QA dataset. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Separate observation from interpretation. A code or measured number belongs to the original record; a sentence explaining that value may be revised, but the revision must be traceable and scientifically warranted.

Transfer Exercise 4: From Field Repair to Release Decision

Take two cases in json sidecars, inheritance, units and machine-readable constraints and build a miniature bilingual review bundle. Include the literal field path, a valid fictional source record, the original explanation, a candidate target explanation, a deliberately faulty alternative and the program or human decision affected. Then create a one-page disposition table: accept the translation, revise the display text, require a dataset correction, request a scientist’s answer, or withhold the material because of privacy. The important skill is knowing where to act. Localization belongs in the display layer when identity and values are fixed; correcting recorded evidence belongs in a governed release with transparent provenance.

Field-Level Evidence Workshop 5: TSV tables, data dictionaries and phenotypic instruments

Tab-separated tables are deceptively approachable because they resemble spreadsheets. In BIDS, a column header can be a machine-read variable, a cell can be a coded response and an associated JSON file can define levels, units and descriptions. Translating every visible string indiscriminately can move a result between conditions or erase what was actually measured.

Use this module as a field dossier, not as a universal substitute for versioned specification text. In each fictional example, the source laboratory has the duty to define what was measured, the data steward controls authorized changes, the translator handles cross-language expression, and an independent analyst checks whether the same decision can be made from the target version. The examples are intentionally illustrative; they are not real participant records, release manifests or clinical advice.

Diagnostic 033: participants.tsv participant_id

Source situation. participant_id associates records with subject folders, so equality and exact spelling matter. In this fictional BIDS dataset, the translator encounters participants.tsv participant_id while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A workbook imports identifiers as numbers and drops leading zeros while translated headers are saved. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Preserve identifiers as strings and do not translate the defined column name. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. Cross-file joins and the archived TSV show intended referential identity. A practical acceptance test is: Perform a full join coverage check before and after export. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Use an independent reviewer who can explain the experimental process rather than merely certify fluency. Give that reviewer the source field, its neighboring fields and a small reproducible sample.

Diagnostic 034: phenotype questionnaire scale

Source situation. A score may derive from a named instrument and a validated language-specific form. In this fictional BIDS dataset, the translator encounters phenotype questionnaire scale while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. The translator replaces the original instrument wording ad hoc and labels scores as equivalent. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Use authorized validated instruments when required, or clearly label unvalidated adaptations. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The instrument manual, license, adaptation study and scoring rules establish boundaries. A practical acceptance test is: Check item mapping, response options, scoring direction and permission. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Treat semantic equivalence as an evidence question. Identify the experiment, provenance or versioned specification that would disprove a tempting but incorrect paraphrase.

Diagnostic 035: response code levels

Source situation. levels in a sidecar data dictionary can map raw codes such as 1 and 2 to specific categories. In this fictional BIDS dataset, the translator encounters response code levels while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. An editor swaps descriptions because one category sounds socially preferable in a target language. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Preserve code-to-level mapping and change only approved display descriptions. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The questionnaire codebook and raw entry form show category meanings. A practical acceptance test is: For every code, ask whether the target-language reader selects the same response. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Record the test at the correct level. A JSON parse is useful for keys and types; it cannot determine whether a cognitive condition has been translated into the wrong construct.

Diagnostic 036: missing value n/a

Source situation. n/a can carry a missing-data convention in tabular records without equaling negative or zero. In this fictional BIDS dataset, the translator encounters missing value n/a while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. The translation calls all n/a responses ‘none’, changing nonresponse into a substantive answer. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Define unavailable, not asked, inapplicable and refused separately when original records support it. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The data dictionary and source collection rules specify missingness. A practical acceptance test is: Count changed missing and nonmissing categories after normalization. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Check portability and accessibility together. A target-language reader should be able to copy a code, locate a file and understand an uncertainty note without relying on hidden English instructions.

Diagnostic 037: age units and derived variables

Source situation. A derived age or time-since-treatment variable needs reference time, units and transformation provenance. In this fictional BIDS dataset, the translator encounters age units and derived variables while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. The target version says months where the numeric values remain in years. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep numeric values and an explicit unit declaration; create a separate view for converted displays. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The processing script and phenotypic sidecar define the calculation. A practical acceptance test is: Recalculate selected records from authorized source values. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Inspect the object as both a scientist and a file parser. The human account must not rely on an assumption that the schema cannot support. If the two descriptions disagree, pause the release and assign the issue to a domain owner.

Diagnostic 038: free-text response translation

Source situation. Participant free text may be a qualitative outcome and a separate translation layer rather than primary data. In this fictional BIDS dataset, the translator encounters free-text response translation while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. The translator replaces the original response so later investigators cannot assess semantic drift. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Retain source text securely and store translated text with language and reviewer provenance. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The research consent, instrument purpose and qualitative coding plan define acceptable handling. A practical acceptance test is: Review paired excerpts without exposing personal details beyond permissions. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Run a reversible test rather than editing the only available copy. Preserve the canonical source, produce an isolated target-language view, and compare what each version causes a reader or program to select.

Diagnostic 039: column description versus variable name

Source situation. A JSON sidecar may describe a custom TSV column while its literal header is used by analysis code. In this fictional BIDS dataset, the translator encounters column description versus variable name while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. An automated localization pass renames the column and breaks the analysis notebook. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Leave column headers stable; translate Description text or attach a localized dictionary. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The schema and actual scripts identify required keys. A practical acceptance test is: Run the same tabular load and verify the same columns, types and joins. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Separate observation from interpretation. A code or measured number belongs to the original record; a sentence explaining that value may be revised, but the revision must be traceable and scientifically warranted.

Diagnostic 040: longitudinal phenotype timing

Source situation. Repeated assessments may be anchored to baseline, intervention or follow-up rather than calendar month. In this fictional BIDS dataset, the translator encounters longitudinal phenotype timing while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. The translation labels ‘week 4’ as four weeks after scanning when the protocol defines four weeks after treatment. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Translate time origin as carefully as the interval and preserve session-level joins. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The protocol calendar and timestamped collection records determine origin. A practical acceptance test is: Reconstruct each sample timeline using the translated definition. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Include the negative case in review. Ask what a researcher could falsely infer when the translation omits a qualification, and whether that inference would affect a published conclusion or participant treatment.

Transfer Exercise 5: From Field Repair to Release Decision

Take two cases in tsv tables, data dictionaries and phenotypic instruments and build a miniature bilingual review bundle. Include the literal field path, a valid fictional source record, the original explanation, a candidate target explanation, a deliberately faulty alternative and the program or human decision affected. Then create a one-page disposition table: accept the translation, revise the display text, require a dataset correction, request a scientist’s answer, or withhold the material because of privacy. The important skill is knowing where to act. Localization belongs in the display layer when identity and values are fixed; correcting recorded evidence belongs in a governed release with transparent provenance.

Field-Level Evidence Workshop 6: Experimental events, stimuli, behavioral timing and HED annotation

Event files connect brain measurements with instructions, stimuli, reactions and analysis regressors. An event label can look like ordinary English while functioning as a factor level. Translation quality is measured by whether a researcher reproduces the same event design, not merely whether trial_type sounds elegant to a reader.

Use this module as a field dossier, not as a universal substitute for versioned specification text. In each fictional example, the source laboratory has the duty to define what was measured, the data steward controls authorized changes, the translator handles cross-language expression, and an independent analyst checks whether the same decision can be made from the target version. The examples are intentionally illustrative; they are not real participant records, release manifests or clinical advice.

Diagnostic 041: events.tsv onset

Source situation. onset is an event time relative to the data acquisition as defined for the modality. In this fictional BIDS dataset, the translator encounters events.tsv onset while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translated note says wall-clock time and invites researchers to compare 15.2 to local time. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Preserve the numeric onset and explain its reference point and units. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The experiment log and modality section establish time alignment. A practical acceptance test is: Plot translated event labels over the actual recording timeline. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Treat semantic equivalence as an evidence question. Identify the experiment, provenance or versioned specification that would disprove a tempting but incorrect paraphrase.

Diagnostic 042: events.tsv duration

Source situation. duration describes event length, and a zero may represent an instantaneous event. In this fictional BIDS dataset, the translator encounters events.tsv duration while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. An editor interprets zero as missing and deletes that event. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep zero durations as recorded and distinguish them from unavailable durations. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The task design and synchronization log show event semantics. A practical acceptance test is: Compare event counts and durations after transformation. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Record the test at the correct level. A JSON parse is useful for keys and types; it cannot determine whether a cognitive condition has been translated into the wrong construct.

Diagnostic 043: trial_type factor values

Source situation. trial_type often labels experimental conditions consumed by statistical models. In this fictional BIDS dataset, the translator encounters trial_type factor values while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translator replaces go/no-go condition tokens in TSV but not the design matrix specification. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Retain factor codes and add human-readable translated definitions in the JSON sidecar or companion. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The task script and contrast definitions establish condition logic. A practical acceptance test is: Rebuild the condition design matrix and compare column identities. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Check portability and accessibility together. A target-language reader should be able to copy a code, locate a file and understand an uncertainty note without relying on hidden English instructions.

Diagnostic 044: response_time sign

Source situation. response_time is an observed delay and may use special entries for absent responses. In this fictional BIDS dataset, the translator encounters response_time sign while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. The translated description treats a negative value as fast performance rather than a coded special case. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Check the study-specific codebook before labeling response failures or excluded trials. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The behavioral parser and quality-control notes govern exceptional values. A practical acceptance test is: Inspect original trials and ensure exclusions remain identical. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Inspect the object as both a scientist and a file parser. The human account must not rely on an assumption that the schema cannot support. If the two descriptions disagree, pause the release and assign the issue to a domain owner.

Diagnostic 045: stim_file references

Source situation. stim_file may point to stimuli paths; images and audio need exact file identity. In this fictional BIDS dataset, the translator encounters stim_file references while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translator substitutes a caption for a stimulus filename and severs the file reference. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep the path and provide translated alt text or description separately. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The archived stimuli directory and experiment script identify presented content. A practical acceptance test is: Open every referenced item and compare checksums. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Run a reversible test rather than editing the only available copy. Preserve the canonical source, produce an isolated target-language view, and compare what each version causes a reader or program to select.

Diagnostic 046: HED tag representation

Source situation. Hierarchical Event Descriptors tags are controlled ontology annotations rather than ordinary prose. In this fictional BIDS dataset, the translator encounters HED tag representation while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translator turns an HED tag path into conversational language in the field parsed by tools. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep HED syntax and vocabulary fixed; provide localized explanations in separate documentation. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The versioned HED schema and event tagging guidelines define allowable tags. A practical acceptance test is: Validate tag structure and test event queries on the dataset. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Separate observation from interpretation. A code or measured number belongs to the original record; a sentence explaining that value may be revised, but the revision must be traceable and scientifically warranted.

Diagnostic 047: instructions versus actual stimuli

Source situation. Participant-facing instructions may differ from analysts’ shorthand names of conditions. In this fictional BIDS dataset, the translator encounters instructions versus actual stimuli while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. An editor translates an instruction but assumes that the recorded condition codes also changed. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep archival distinction between what participants saw and what data analysts coded. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. Stimulus presentation files and trial logs are the primary evidence. A practical acceptance test is: Run a bilingual stimulus audit against condition assignment. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Include the negative case in review. Ask what a researcher could falsely infer when the translation omits a qualification, and whether that inference would affect a published conclusion or participant treatment.

Diagnostic 048: event clock drift and triggers

Source situation. Synchronization can involve triggers, TTL pulses and device latency corrections. In this fictional BIDS dataset, the translator encounters event clock drift and triggers while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translated report calls a corrected timestamp the raw event time or discards its correction note. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Preserve the timing provenance and describe the correction in clear target-language prose. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. Hardware logs and preprocessing code establish the correction method. A practical acceptance test is: Compare pre-correction and post-correction event plots. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Use an independent reviewer who can explain the experimental process rather than merely certify fluency. Give that reviewer the source field, its neighboring fields and a small reproducible sample.

Transfer Exercise 6: From Field Repair to Release Decision

Take two cases in experimental events, stimuli, behavioral timing and hed annotation and build a miniature bilingual review bundle. Include the literal field path, a valid fictional source record, the original explanation, a candidate target explanation, a deliberately faulty alternative and the program or human decision affected. Then create a one-page disposition table: accept the translation, revise the display text, require a dataset correction, request a scientist’s answer, or withhold the material because of privacy. The important skill is knowing where to act. Localization belongs in the display layer when identity and values are fixed; correcting recorded evidence belongs in a governed release with transparent provenance.

Field-Level Evidence Workshop 7: Structural and functional MRI terminology without altered acquisition evidence

Imaging words can refer to contrast, sequence, acquisition mechanism or postprocessed display. A translated caption must not convert a protocol label into a clinical conclusion. This module keeps the BIDS MRI data model separate from general medical-image translation, which remains its own subject elsewhere in the site.

Use this module as a field dossier, not as a universal substitute for versioned specification text. In each fictional example, the source laboratory has the duty to define what was measured, the data steward controls authorized changes, the translator handles cross-language expression, and an independent analyst checks whether the same decision can be made from the target version. The examples are intentionally illustrative; they are not real participant records, release manifests or clinical advice.

Diagnostic 049: T1w versus T2w

Source situation. T1w and T2w label image weighting, not a diagnosis or whether tissue is healthy. In this fictional BIDS dataset, the translator encounters T1w versus T2w while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A multilingual figure legend calls T2w a pathology scan and implies abnormalities are present. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Retain suffixes and describe weighting without making clinical claims. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The scanner series description and protocol identify contrast. A practical acceptance test is: Match images to sequence metadata rather than visual guesses. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Record the test at the correct level. A JSON parse is useful for keys and types; it cannot determine whether a cognitive condition has been translated into the wrong construct.

Diagnostic 050: FLAIR acquisition

Source situation. FLAIR is a fluid-attenuated inversion recovery MRI contrast label. In this fictional BIDS dataset, the translator encounters FLAIR acquisition while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. An editor translates the acronym into a literal phrase without identifying it as a specific technique. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep the recognized sequence label and provide a plain-language expansion. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The acquisition protocol documents sequence parameters. A practical acceptance test is: Verify the reader knows the label is a scan type, not an interpretation. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Check portability and accessibility together. A target-language reader should be able to copy a code, locate a file and understand an uncertainty note without relying on hidden English instructions.

Diagnostic 051: BOLD signals

Source situation. BOLD refers to a blood-oxygenation-level-dependent contrast used in functional MRI. In this fictional BIDS dataset, the translator encounters BOLD signals while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translated article equates greater BOLD response with more thoughts or direct neuronal firing. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Describe BOLD as an indirect hemodynamic signal with analytic qualifications. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The study methods and fMRI model establish what was estimated. A practical acceptance test is: Have an analyst compare the target claim to the actual statistical contrast. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Inspect the object as both a scientist and a file parser. The human account must not rely on an assumption that the schema cannot support. If the two descriptions disagree, pause the release and assign the issue to a domain owner.

Diagnostic 052: multi-echo data

Source situation. echo entities and EchoTime values distinguish multiple measurements from the same acquisition. In this fictional BIDS dataset, the translator encounters multi-echo data while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translation collapses echo-1 and echo-2 into one generic image caption. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Preserve separate echo indices and timing arrays. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The acquisition sequence and collection definitions show linkage. A practical acceptance test is: Check that all echoes are still discovered and grouped correctly. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Run a reversible test rather than editing the only available copy. Preserve the canonical source, produce an isolated target-language view, and compare what each version causes a reader or program to select.

Diagnostic 053: resting-state instructions

Source situation. A task-rest file says what acquisition category was used but not necessarily what the participant did with eyes. In this fictional BIDS dataset, the translator encounters resting-state instructions while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translation says eyes closed even though study subjects were instructed to fixate a cross. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Use actual participant instruction text and distinguish it from resting-state label. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The instruction script and protocol establish eyes-open/closed status. A practical acceptance test is: Compare translated documentation to a sample run’s stimuli. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Separate observation from interpretation. A code or measured number belongs to the original record; a sentence explaining that value may be revised, but the revision must be traceable and scientifically warranted.

Diagnostic 054: slice orientation

Source situation. Slice directions belong to image axes and acquisition geometry; colloquial up/down may mislead. In this fictional BIDS dataset, the translator encounters slice orientation while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A bilingual illustration mirrors a scan because the translator assumes right means right of the page. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Preserve orientation metadata and localize labels with anatomical frame references. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The NIfTI affine and acquisition orientation govern geometry. A practical acceptance test is: Review a laterality landmark against the coordinate frame. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Include the negative case in review. Ask what a researcher could falsely infer when the translation omits a qualification, and whether that inference would affect a published conclusion or participant treatment.

Diagnostic 055: brain mask derivatives

Source situation. A brain mask is an analysis product with binary or probabilistic values, not a categorical anatomical diagnosis. In this fictional BIDS dataset, the translator encounters brain mask derivatives while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A caption calls the whole mask abnormal brain tissue after misreading label 1. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Describe what the segmentation covers and how values were produced. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The generating pipeline, threshold settings and output sidecar define meaning. A practical acceptance test is: Overlay the mask and check value descriptions with the producer. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Use an independent reviewer who can explain the experimental process rather than merely certify fluency. Give that reviewer the source field, its neighboring fields and a small reproducible sample.

Diagnostic 056: quality metrics on scans

Source situation. MRI quality indices quantify specific artefacts or signal characteristics, not total subject reliability. In this fictional BIDS dataset, the translator encounters quality metrics on scans while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translated dashboard turns an outlier flag into an automatic declaration that a participant’s data are invalid. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep metric formulas, thresholds and review status distinct. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The quality-control protocol and software version show acceptance criteria. A practical acceptance test is: Check whether the translated advice would trigger unapproved exclusions. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Treat semantic equivalence as an evidence question. Identify the experiment, provenance or versioned specification that would disprove a tempting but incorrect paraphrase.

Transfer Exercise 7: From Field Repair to Release Decision

Take two cases in structural and functional mri terminology without altered acquisition evidence and build a miniature bilingual review bundle. Include the literal field path, a valid fictional source record, the original explanation, a candidate target explanation, a deliberately faulty alternative and the program or human decision affected. Then create a one-page disposition table: accept the translation, revise the display text, require a dataset correction, request a scientist’s answer, or withhold the material because of privacy. The important skill is knowing where to act. Localization belongs in the display layer when identity and values are fixed; correcting recorded evidence belongs in a governed release with transparent provenance.

Field-Level Evidence Workshop 8: Diffusion, fieldmaps, perfusion and cross-modal timing

BIDS supports several MRI-related representations that have similar-looking fields but different physical interpretations. The translator must protect measurement units, directions, spatial mapping, and the boundary between instrument acquisition and later calculation. Where a definition has changed in a particular specification version, the recorded dataset and declared edition determine how it should be reviewed.

Use this module as a field dossier, not as a universal substitute for versioned specification text. In each fictional example, the source laboratory has the duty to define what was measured, the data steward controls authorized changes, the translator handles cross-language expression, and an independent analyst checks whether the same decision can be made from the target version. The examples are intentionally illustrative; they are not real participant records, release manifests or clinical advice.

Diagnostic 057: DWI bval files

Source situation. b-values describe diffusion-weighting parameters, usually in s/mm². In this fictional BIDS dataset, the translator encounters DWI bval files while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. An editor translates s/mm² as mm²/s, accidentally reversing the physical dimension. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Freeze data and preserve the precise unit in explanatory prose. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The diffusion protocol and BIDS diffusion rules specify b-value interpretation. A practical acceptance test is: Check count and order against acquired volumes. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Check portability and accessibility together. A target-language reader should be able to copy a code, locate a file and understand an uncertainty note without relying on hidden English instructions.

Diagnostic 058: DWI bvec directions

Source situation. b-vectors represent gradient directions in the coordinate convention relevant to the data. In this fictional BIDS dataset, the translator encounters DWI bvec directions while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translated note treats the vector components as east, north and up without reference to image axes. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Explain the coordinate basis and do not rotate numerical vectors for language reasons. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The conversion and gradient-rotation provenance are essential evidence. A practical acceptance test is: Visualize directions relative to known image orientation. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Inspect the object as both a scientist and a file parser. The human account must not rely on an assumption that the schema cannot support. If the two descriptions disagree, pause the release and assign the issue to a domain owner.

Diagnostic 059: fieldmap association

Source situation. Fieldmap magnitude, phase and acquisition information support geometric distortion correction. In this fictional BIDS dataset, the translator encounters fieldmap association while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translation calls phase and magnitude interchangeable channels and loses a suffix distinction. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Retain suffix, units and IntendedFor mapping; explain their different roles. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. Source scanner metadata and fieldmap convention establish meaning. A practical acceptance test is: Check distortion-correction inputs and outputs are unchanged. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Run a reversible test rather than editing the only available copy. Preserve the canonical source, produce an isolated target-language view, and compare what each version causes a reader or program to select.

Diagnostic 060: TotalReadoutTime

Source situation. TotalReadoutTime has a specific definition in echo-planar imaging metadata. In this fictional BIDS dataset, the translator encounters TotalReadoutTime while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A bilingual glossary replaces it with total scan duration. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Describe echo-planar readout rather than the total time participant spent in scanner. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The sequence parameter calculation shows the source quantity. A practical acceptance test is: Recompute a sample from acquisition metadata. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Separate observation from interpretation. A code or measured number belongs to the original record; a sentence explaining that value may be revised, but the revision must be traceable and scientifically warranted.

Diagnostic 061: ASL labeling duration

Source situation. Arterial spin labeling uses labeling and post-labeling timing parameters with specific physiological roles. In this fictional BIDS dataset, the translator encounters ASL labeling duration while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translation merges labeling duration with post-labeling delay. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep these fields distinct and explain each time interval with a timeline. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The ASL acquisition method and protocol are authoritative. A practical acceptance test is: Draw the labeled acquisition timeline and compare field assignments. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Include the negative case in review. Ask what a researcher could falsely infer when the translation omits a qualification, and whether that inference would affect a published conclusion or participant treatment.

Diagnostic 062: blood flow units

Source situation. A perfusion output may be reported with flow units per tissue mass and time. In this fictional BIDS dataset, the translator encounters blood flow units while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A target summary says raw intensity without stating the normalized physiological quantity. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Preserve unit, normalization, preprocessing and measurement status. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The quantification pipeline and model definitions explain the output. A practical acceptance test is: Reconstruct one sample calculation with the original units. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Use an independent reviewer who can explain the experimental process rather than merely certify fluency. Give that reviewer the source field, its neighboring fields and a small reproducible sample.

Diagnostic 063: physiological channels and sampling

Source situation. Respiration and cardiac traces may be sampled independently from imaging volumes. In this fictional BIDS dataset, the translator encounters physiological channels and sampling while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. An editor changes sampling frequency units or calls the stream frame rate. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep sampling frequency, start time and column mapping explicit. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The device export and clock synchronization record anchor timing. A practical acceptance test is: Plot the synchronized trace against the scan. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Treat semantic equivalence as an evidence question. Identify the experiment, provenance or versioned specification that would disprove a tempting but incorrect paraphrase.

Diagnostic 064: cross-modality correspondence

Source situation. A paper may match structural, functional and diffusion images from the same participant and session. In this fictional BIDS dataset, the translator encounters cross-modality correspondence while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A localized index matches similar filenames across different participants. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Preserve subject/session keys, IntendedFor links and coordinate-space provenance. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The dataset manifest and transformations define legitimate correspondence. A practical acceptance test is: Verify linked image transforms and IDs for random samples. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Record the test at the correct level. A JSON parse is useful for keys and types; it cannot determine whether a cognitive condition has been translated into the wrong construct.

Transfer Exercise 8: From Field Repair to Release Decision

Take two cases in diffusion, fieldmaps, perfusion and cross-modal timing and build a miniature bilingual review bundle. Include the literal field path, a valid fictional source record, the original explanation, a candidate target explanation, a deliberately faulty alternative and the program or human decision affected. Then create a one-page disposition table: accept the translation, revise the display text, require a dataset correction, request a scientist’s answer, or withhold the material because of privacy. The important skill is knowing where to act. Localization belongs in the display layer when identity and values are fixed; correcting recorded evidence belongs in a governed release with transparent provenance.

Field-Level Evidence Workshop 9: EEG, MEG, iEEG and electrophysiology channel semantics

Electrical recordings contain channel labels, measurement units, sensor geometry, reference schemes and event triggers. Their files may be binary or instrument-specific and are not translated as prose. A reliable multilingual data dictionary should explain what a channel measures without changing the signal identity, reference chain or analyst selection.

Use this module as a field dossier, not as a universal substitute for versioned specification text. In each fictional example, the source laboratory has the duty to define what was measured, the data steward controls authorized changes, the translator handles cross-language expression, and an independent analyst checks whether the same decision can be made from the target version. The examples are intentionally illustrative; they are not real participant records, release manifests or clinical advice.

Diagnostic 065: eeg.json PowerLineFrequency

Source situation. PowerLineFrequency describes nominal mains interference frequency for an EEG recording. In this fictional BIDS dataset, the translator encounters eeg.json PowerLineFrequency while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. The translated documentation treats it as the EEG sampling rate. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep the field and explain difference from SamplingFrequency. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. Acquisition metadata and channel files distinguish frequencies. A practical acceptance test is: Inspect the power spectrum and JSON fields. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Inspect the object as both a scientist and a file parser. The human account must not rely on an assumption that the schema cannot support. If the two descriptions disagree, pause the release and assign the issue to a domain owner.

Diagnostic 066: EEG SamplingFrequency

Source situation. SamplingFrequency is the signal acquisition rate in hertz. In this fictional BIDS dataset, the translator encounters EEG SamplingFrequency while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translation converts 500 Hz into 500 seconds or rounds to 0.5 kHz in the raw field. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep numeric field stable; offer optional human-readable conversion. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The recording header and protocol define sample rate. A practical acceptance test is: Compare duration and sample count to the expected data array. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Run a reversible test rather than editing the only available copy. Preserve the canonical source, produce an isolated target-language view, and compare what each version causes a reader or program to select.

Diagnostic 067: channels.tsv type

Source situation. Channel type describes sensor data categories, and codes may be standardized or modality-dependent. In this fictional BIDS dataset, the translator encounters channels.tsv type while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. An editor translates a code EEG into a natural-language word in the typed column. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Preserve machine category codes and explain them in the sidecar. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The channels schema and source equipment configuration are authority. A practical acceptance test is: Validate allowed values and ensure signal import works. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Separate observation from interpretation. A code or measured number belongs to the original record; a sentence explaining that value may be revised, but the revision must be traceable and scientifically warranted.

Diagnostic 068: channels.tsv status

Source situation. Bad-channel annotations report processing decisions or signal limitations rather than a clinical outcome. In this fictional BIDS dataset, the translator encounters channels.tsv status while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translator renders bad as dangerous to the participant. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Clarify whether status means noisy, excluded or affected by recording artefact. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. Channel quality logs and annotations show grounds. A practical acceptance test is: Test that software excludes the same channels. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Include the negative case in review. Ask what a researcher could falsely infer when the translation omits a qualification, and whether that inference would affect a published conclusion or participant treatment.

Diagnostic 069: electrodes.tsv spatial units

Source situation. Electrode positions depend on a coordinate system and an associated units declaration. In this fictional BIDS dataset, the translator encounters electrodes.tsv spatial units while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A multilingual image shifts a point by treating millimetres as metres. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Preserve coordinates and unit labels and provide a separately generated display. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. Coordsystem JSON and digitization record establish frame. A practical acceptance test is: Overlay electrodes on the correct head model. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Use an independent reviewer who can explain the experimental process rather than merely certify fluency. Give that reviewer the source field, its neighboring fields and a small reproducible sample.

Diagnostic 070: EEG reference

Source situation. An EEG reference scheme changes voltage interpretation and is not an optional caption. In this fictional BIDS dataset, the translator encounters EEG reference while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. The translated methods sentence implies average reference when the signal used mastoids. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep the actual reference text and link it to preprocessing steps. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. Recording configuration and processing code show reference. A practical acceptance test is: Compare rerendered traces and interpretation under correct reference. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Treat semantic equivalence as an evidence question. Identify the experiment, provenance or versioned specification that would disprove a tempting but incorrect paraphrase.

Diagnostic 071: MEG sensor system

Source situation. MEG measurements may use magnetometers and gradiometers with different units and systems. In this fictional BIDS dataset, the translator encounters MEG sensor system while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. An editor calls every sensor a magnetometer and hides differences. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Translate instrument descriptions precisely and preserve channel type codes. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. Instrument documentation and channel metadata establish sensor type. A practical acceptance test is: Plot channel types separately and check expected counts. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Record the test at the correct level. A JSON parse is useful for keys and types; it cannot determine whether a cognitive condition has been translated into the wrong construct.

Diagnostic 072: iEEG electrode laterality

Source situation. Intracranial EEG electrode labels and coordinates have safety-critical anatomical meaning. In this fictional BIDS dataset, the translator encounters iEEG electrode laterality while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translator switches left and right when presenting a mirrored figure. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep recorded labels and use coordinate-based laterality review. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. Implant record and imaging registration provide evidence under authorized access. A practical acceptance test is: Independently check left/right designation on an orientation-marked image. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Check portability and accessibility together. A target-language reader should be able to copy a code, locate a file and understand an uncertainty note without relying on hidden English instructions.

Transfer Exercise 9: From Field Repair to Release Decision

Take two cases in eeg, meg, ieeg and electrophysiology channel semantics and build a miniature bilingual review bundle. Include the literal field path, a valid fictional source record, the original explanation, a candidate target explanation, a deliberately faulty alternative and the program or human decision affected. Then create a one-page disposition table: accept the translation, revise the display text, require a dataset correction, request a scientist’s answer, or withhold the material because of privacy. The important skill is knowing where to act. Localization belongs in the display layer when identity and values are fixed; correcting recorded evidence belongs in a governed release with transparent provenance.

Field-Level Evidence Workshop 10: PET, NIRS, microscopy, spectroscopy and other modalities

A multimodal dataset requires the translator to understand which terms are measurement conditions, which are biological interpretations and which are metadata about instruments. BIDS version 1.11.2 lists more modalities than the older MRI-only picture. These case studies teach cross-modality discipline rather than an exhaustive replacement for each specialized standard.

Use this module as a field dossier, not as a universal substitute for versioned specification text. In each fictional example, the source laboratory has the duty to define what was measured, the data steward controls authorized changes, the translator handles cross-language expression, and an independent analyst checks whether the same decision can be made from the target version. The examples are intentionally illustrative; they are not real participant records, release manifests or clinical advice.

Diagnostic 073: PET tracer names

Source situation. Tracer identity and radionuclide information are distinct from a generic label of radioactive scan. In this fictional BIDS dataset, the translator encounters PET tracer names while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. The translated study description substitutes the wrong tracer shorthand. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Retain the exact tracer identity and its applicable metadata. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The PET protocol, radiochemistry documentation and acquisition sidecars confirm tracer. A practical acceptance test is: Compare the administered tracer record with analyzed series. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Run a reversible test rather than editing the only available copy. Preserve the canonical source, produce an isolated target-language view, and compare what each version causes a reader or program to select.

Diagnostic 074: PET injected dose

Source situation. Injected dose is a recorded quantity with a prescribed unit and reference time. In this fictional BIDS dataset, the translator encounters PET injected dose while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. An editor confuses megabecquerels with milligrams and creates a dosing claim. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep numeric source unchanged; explain radioactivity separately from mass. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The radiopharmaceutical administration log establishes measure. A practical acceptance test is: Review unit conversions with qualified domain experts. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Separate observation from interpretation. A code or measured number belongs to the original record; a sentence explaining that value may be revised, but the revision must be traceable and scientifically warranted.

Diagnostic 075: PET time-activity curves

Source situation. Frames describe temporal measurement intervals and can affect modeling. In this fictional BIDS dataset, the translator encounters PET time-activity curves while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translation labels frame midpoint as duration. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Preserve frame start, length and temporal origin separately. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The frame timing file and kinetic model define interpretation. A practical acceptance test is: Reconstruct the frame boundaries on a timeline. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Include the negative case in review. Ask what a researcher could falsely infer when the translation omits a qualification, and whether that inference would affect a published conclusion or participant treatment.

Diagnostic 076: NIRS source-detector geometry

Source situation. Near-infrared spectroscopy arrays depend on source-detector pairing and optode placements. In this fictional BIDS dataset, the translator encounters NIRS source-detector geometry while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. An editor renames channel labels from their translated anatomical descriptions. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep numeric and channel IDs fixed while translating approved display names. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. Optode layout, measurement list and coordsystem metadata determine meaning. A practical acceptance test is: Recompute sample source-detector pairing. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Use an independent reviewer who can explain the experimental process rather than merely certify fluency. Give that reviewer the source field, its neighboring fields and a small reproducible sample.

Diagnostic 077: microscopy sample labels

Source situation. A sample identity remains tied to tissue preparation and instrument metadata. In this fictional BIDS dataset, the translator encounters microscopy sample labels while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translation changes a sample label into an assumed tissue diagnosis. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Use approved specimen descriptors and preserve sample keys. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. Collection protocol and sample manifest provide source facts. A practical acceptance test is: Check each image is associated with the same sample. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Treat semantic equivalence as an evidence question. Identify the experiment, provenance or versioned specification that would disprove a tempting but incorrect paraphrase.

Diagnostic 078: MRS metabolite names

Source situation. Magnetic resonance spectroscopy measures signals that can be modeled as metabolite-related quantities. In this fictional BIDS dataset, the translator encounters MRS metabolite names while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translated chart turns a modeled metabolite estimate into a definitive diagnostic claim. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep metabolite notation and uncertainty; distinguish fit quality from clinical interpretation. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. Acquisition and fitting pipeline records define estimates. A practical acceptance test is: Compare display labels to the actual fitted metabolites. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Record the test at the correct level. A JSON parse is useful for keys and types; it cannot determine whether a cognitive condition has been translated into the wrong construct.

Diagnostic 079: motion capture coordinate data

Source situation. Motion capture tracks marker positions with coordinate frames and time origins. In this fictional BIDS dataset, the translator encounters motion capture coordinate data while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. An editor rewords positive axis as leftward without considering chosen convention. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Describe the frame; leave positions and coordinate transforms untouched. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. Camera calibration and marker-mapping files establish geometry. A practical acceptance test is: Animate a known movement and check laterality. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Check portability and accessibility together. A target-language reader should be able to copy a code, locate a file and understand an uncertainty note without relying on hidden English instructions.

Diagnostic 080: EMG muscle recording

Source situation. Electromyography channels measure muscle electrical activity, not simply movement speed. In this fictional BIDS dataset, the translator encounters EMG muscle recording while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translation labels EMG amplitude as muscle force in newtons. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Preserve the recorded electrical measure and distinguish inference from observed signal. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. Acquisition montage and signal-processing record establish modality. A practical acceptance test is: Check signal units and captions against a raw trace. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Inspect the object as both a scientist and a file parser. The human account must not rely on an assumption that the schema cannot support. If the two descriptions disagree, pause the release and assign the issue to a domain owner.

Transfer Exercise 10: From Field Repair to Release Decision

Take two cases in pet, nirs, microscopy, spectroscopy and other modalities and build a miniature bilingual review bundle. Include the literal field path, a valid fictional source record, the original explanation, a candidate target explanation, a deliberately faulty alternative and the program or human decision affected. Then create a one-page disposition table: accept the translation, revise the display text, require a dataset correction, request a scientist’s answer, or withhold the material because of privacy. The important skill is knowing where to act. Localization belongs in the display layer when identity and values are fixed; correcting recorded evidence belongs in a governed release with transparent provenance.

Field-Level Evidence Workshop 11: Derivatives, transforms, atlases, quality assurance and provenance

A derivative is an output of a process, not another raw measurement. Translation must preserve the transformation chain, source file identity, tool versions and interpretation boundaries. Especially where templates and atlases are used, apparently familiar anatomical words have spatial and statistical definitions that should not be flattened.

Use this module as a field dossier, not as a universal substitute for versioned specification text. In each fictional example, the source laboratory has the duty to define what was measured, the data steward controls authorized changes, the translator handles cross-language expression, and an independent analyst checks whether the same decision can be made from the target version. The examples are intentionally illustrative; they are not real participant records, release manifests or clinical advice.

Diagnostic 081: RawSources list

Source situation. RawSources can document the raw inputs from which a derivative was made. In this fictional BIDS dataset, the translator encounters RawSources list while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translator changes filenames into readable titles, destroying provenance joins. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep exact paths and add labeled descriptions beside references. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The derivative metadata and raw manifest show input lineage. A practical acceptance test is: Resolve every source and compare content hashes. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Separate observation from interpretation. A code or measured number belongs to the original record; a sentence explaining that value may be revised, but the revision must be traceable and scientifically warranted.

Diagnostic 082: SpatialReference metadata

Source situation. SpatialReference indicates a reference space or image where applicable. In this fictional BIDS dataset, the translator encounters SpatialReference metadata while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. An editor substitutes a translated colloquial name for a standardized template identifier. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep the standard space code and translate its explanation. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. Template registry, registration command and output provenance identify coordinate space. A practical acceptance test is: Overlay known landmarks using declared transform. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Include the negative case in review. Ask what a researcher could falsely infer when the translation omits a qualification, and whether that inference would affect a published conclusion or participant treatment.

Diagnostic 083: resampling and interpolation

Source situation. Resampling is a computational transform that can change spatial grid properties. In this fictional BIDS dataset, the translator encounters resampling and interpolation while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A target description claims the image was acquired at the new resolution. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. State that resolution is derivative and document interpolation method. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. Processing logs and before/after headers show resampling. A practical acceptance test is: Compare voxel dimensions and transform chain. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Use an independent reviewer who can explain the experimental process rather than merely certify fluency. Give that reviewer the source field, its neighboring fields and a small reproducible sample.

Diagnostic 084: atlas label indices

Source situation. Atlas label values often map integers to named regions in a separate table. In this fictional BIDS dataset, the translator encounters atlas label indices while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translator reorders region names alphabetically without keeping numeric mapping. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Preserve label integer-to-region mapping and use reviewed regional names. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. Atlas reference, label TSV and documentation define correspondence. A practical acceptance test is: Look up random voxels and confirm region IDs agree. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Treat semantic equivalence as an evidence question. Identify the experiment, provenance or versioned specification that would disprove a tempting but incorrect paraphrase.

Diagnostic 085: probability maps

Source situation. A probabilistic atlas or mask encodes fractional membership rather than categorical truth. In this fictional BIDS dataset, the translator encounters probability maps while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translation calls 0.6 certain membership. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Explain probability-like representation and thresholding if used. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. Atlas methodology and threshold definition establish meaning. A practical acceptance test is: Sample voxel values and verify thresholds unchanged. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Record the test at the correct level. A JSON parse is useful for keys and types; it cannot determine whether a cognitive condition has been translated into the wrong construct.

Diagnostic 086: motion confounds

Source situation. Framewise displacement and other confounds are measures derived by particular algorithms. In this fictional BIDS dataset, the translator encounters motion confounds while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. An editor calls a confound score proof of participant noncompliance. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Identify metric derivation, units, caveats and pipeline. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. Methods file and software version determine computation. A practical acceptance test is: Compare source metrics to the localized QA table. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Check portability and accessibility together. A target-language reader should be able to copy a code, locate a file and understand an uncertainty note without relying on hidden English instructions.

Diagnostic 087: QC exclusion flags

Source situation. Quality-control labels often reflect predefined or manual acceptance procedures. In this fictional BIDS dataset, the translator encounters QC exclusion flags while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A translation reclassifies borderline as rejected based on emotional connotation. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Preserve the recorded review state and publish the rule used. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The reviewer log and exclusion policy provide decision authority. A practical acceptance test is: Recompute participant counts under both language views. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Inspect the object as both a scientist and a file parser. The human account must not rely on an assumption that the schema cannot support. If the two descriptions disagree, pause the release and assign the issue to a domain owner.

Diagnostic 088: pipeline software versions

Source situation. Reproducibility depends on exact tool identity, container image and parameter settings. In this fictional BIDS dataset, the translator encounters pipeline software versions while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. An editor generalizes a versioned command into an unnamed latest software tool. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep release strings and parameters and translate methodological explanation only. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. Execution manifest and run log link files to the producing process. A practical acceptance test is: Execute or inspect a representative pipeline to verify the same inputs. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Run a reversible test rather than editing the only available copy. Preserve the canonical source, produce an isolated target-language view, and compare what each version causes a reader or program to select.

Transfer Exercise 11: From Field Repair to Release Decision

Take two cases in derivatives, transforms, atlases, quality assurance and provenance and build a miniature bilingual review bundle. Include the literal field path, a valid fictional source record, the original explanation, a candidate target explanation, a deliberately faulty alternative and the program or human decision affected. Then create a one-page disposition table: accept the translation, revise the display text, require a dataset correction, request a scientist’s answer, or withhold the material because of privacy. The important skill is knowing where to act. Localization belongs in the display layer when identity and values are fixed; correcting recorded evidence belongs in a governed release with transparent provenance.

Field-Level Evidence Workshop 12: Localization engineering, public display, accessibility and release governance

Translation in research data is a release-management job. The safe deliverable is usually a localized presentation and controlled annotation layer tied to an immutable, auditable dataset. A team should not conflate the absence of a parser error with semantic equivalence, nor assume privacy is guaranteed by the presence of a pseudonym.

Use this module as a field dossier, not as a universal substitute for versioned specification text. In each fictional example, the source laboratory has the duty to define what was measured, the data steward controls authorized changes, the translator handles cross-language expression, and an independent analyst checks whether the same decision can be made from the target version. The examples are intentionally illustrative; they are not real participant records, release manifests or clinical advice.

Diagnostic 089: localized documentation mirror

Source situation. Researchers want a translated README without losing the canonical data package. In this fictional BIDS dataset, the translator encounters localized documentation mirror while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A web exporter overwrites the original README and removes methodological qualifications. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Keep a versioned localized document linked to the original source record. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The approved documentation release provides the baseline. A practical acceptance test is: Diff both documents and check all technical numbers. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Include the negative case in review. Ask what a researcher could falsely infer when the translation omits a qualification, and whether that inference would affect a published conclusion or participant treatment.

Diagnostic 090: multilingual metadata catalog

Source situation. A discovery portal needs translated dataset summaries for readers in several languages. In this fictional BIDS dataset, the translator encounters multilingual metadata catalog while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. An index uses a translated project description as a new DOI-level identity. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Use one persistent record with language-tagged display labels and canonical links. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The repository identifier and metadata API determine authoritative record. A practical acceptance test is: Resolve each language view to the same deposit. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Use an independent reviewer who can explain the experimental process rather than merely certify fluency. Give that reviewer the source field, its neighboring fields and a small reproducible sample.

Diagnostic 091: HTML language tagging

Source situation. Localized descriptions should signal language to browsers and assistive technologies. In this fictional BIDS dataset, the translator encounters HTML language tagging while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. An editor leaves target-language paragraphs tagged as English. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Apply appropriate BCP 47 language tags to presentation content. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The publication locale and translated content determine tag choice. A practical acceptance test is: Test screen reader language switching and search indexing. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Treat semantic equivalence as an evidence question. Identify the experiment, provenance or versioned specification that would disprove a tempting but incorrect paraphrase.

Diagnostic 092: text expansion in dashboards

Source situation. Technical labels can be longer in a target language than in an English control panel. In this fictional BIDS dataset, the translator encounters text expansion in dashboards while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A UI truncates the difference between recommended and required metadata. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Design responsive labels and explain constraints without shrinking text indiscriminately. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The interface string inventory records available space and semantic priority. A practical acceptance test is: Inspect mobile, zoomed and keyboard navigation states. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Record the test at the correct level. A JSON parse is useful for keys and types; it cannot determine whether a cognitive condition has been translated into the wrong construct.

Diagnostic 093: right-to-left visual labels

Source situation. A right-to-left interface can reorder identifiers, dates and filenames visually. In this fictional BIDS dataset, the translator encounters right-to-left visual labels while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A code token appears backwards or a numeric sequence is visually permuted. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Isolate literal code strings and review bidirectional text rendering separately. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The raw code values and interface test cases establish expected order. A practical acceptance test is: Copy displayed identifiers and compare byte-for-byte. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Check portability and accessibility together. A target-language reader should be able to copy a code, locate a file and understand an uncertainty note without relying on hidden English instructions.

Diagnostic 094: machine translation confidentiality

Source situation. Automatic translation can expose sensitive instrument notes or participant descriptions. In this fictional BIDS dataset, the translator encounters machine translation confidentiality while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A tool uploads protected free text to a provider not authorized by the research data agreement. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Classify data before processing and use only approved environments with appropriate controls. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. Data governance agreements define allowed processing routes. A practical acceptance test is: Inspect transfer logs and vendor retention settings. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Inspect the object as both a scientist and a file parser. The human account must not rely on an assumption that the schema cannot support. If the two descriptions disagree, pause the release and assign the issue to a domain owner.

Diagnostic 095: translation memory contamination

Source situation. Reuse is efficient, but a generic ‘session’ translation can erase BIDS-specific meaning. In this fictional BIDS dataset, the translator encounters translation memory contamination while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A termbase imports clinical appointment wording into all session labels. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Record sense, context, domain and exception in terminology entries. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The BIDS definitions and approved project glossary justify term choice. A practical acceptance test is: Test candidate translations in a full method paragraph, not isolation. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Run a reversible test rather than editing the only available copy. Preserve the canonical source, produce an isolated target-language view, and compare what each version causes a reader or program to select.

Diagnostic 096: release rollback and audit

Source situation. Corrections must preserve source, version, and reason for a change across languages. In this fictional BIDS dataset, the translator encounters release rollback and audit while preparing a second-language explanation for a colleague who did not collect the data. The source string seems ordinary until it is read together with its file, its type and its consumer. Establish whether the target wording is a display label, a structural token, a measured quantity or a governed statement. Then write down what the reader is supposed to infer and what the analysis software is expected to do. This prevents a normal editorial choice from becoming an undocumented transformation of experimental evidence.

A plausible failure. A localized release quietly changes translated inclusion criteria after the original publication. The danger is not simply that one term sounds awkward. A scientist may build the wrong participant cohort, align an event with the wrong scan, describe an experimental intervention incorrectly or rerun an analysis with altered inputs. A file can also remain syntactically valid while its claims become false. Before accepting the change, identify the downstream decision the string influences and the smallest fact needed to reject the bad reading. If the source itself is ambiguous, record that ambiguity and seek the investigator’s clarification rather than selecting the most elegant guess.

Controlled repair. Log every changed sentence, reviewer, reason and affected dataset version. The working translation should preserve the canonical source and keep any new human-readable interpretation in a clearly marked presentation layer. Record the source path, original text or value, target version, language, rationale, reviewer, status and expected invariants. Where a change really does require altering machine-readable metadata, route it through the separately approved dataset-maintenance process and issue a new traceable release; do not disguise it as language editing. This division of labor keeps the translated account understandable while protecting reproducibility and data governance.

Evidence and diagnostic. The issue tracker and content hash identify what changed. A practical acceptance test is: Compare versioned localized exports and ensure rollback restores the matching source. Ask a second reader to locate the field without being coached, explain its scope in their own words and predict what would be different if the proposed mistranslation were used. Record a pass only when both technical behavior and scientific interpretation agree; otherwise keep the case open with a named owner. The final lesson generalizes to other research datasets: translate the reader’s route to a fact, not the identity or magnitude of the fact itself. Separate observation from interpretation. A code or measured number belongs to the original record; a sentence explaining that value may be revised, but the revision must be traceable and scientifically warranted.

Transfer Exercise 12: From Field Repair to Release Decision

Take two cases in localization engineering, public display, accessibility and release governance and build a miniature bilingual review bundle. Include the literal field path, a valid fictional source record, the original explanation, a candidate target explanation, a deliberately faulty alternative and the program or human decision affected. Then create a one-page disposition table: accept the translation, revise the display text, require a dataset correction, request a scientist’s answer, or withhold the material because of privacy. The important skill is knowing where to act. Localization belongs in the display layer when identity and values are fixed; correcting recorded evidence belongs in a governed release with transparent provenance.

Integrated Worked Example: A Cross-Language Research Deposit

Imagine a fictional longitudinal study with two groups of volunteers, three sessions, an MRI task, EEG recordings and questionnaires. The English-language package includes dataset_description.json, participants.tsv with a coded group variable, participants.json describing its levels, two func folders of BOLD images with events files, EEG channels tables, a sourcedata area held under additional access control, and one processing derivative. A translated repository landing page is requested for several research collaborators. The safe project plan begins by checking permissions, fixing a source release and deciding that the scientific package will remain unchanged while human-readable documentation and some approved descriptions receive localized counterparts.

An inexperienced editor first translates filenames, changes coded condition labels inside events.tsv, substitutes a target-language word for JSON true and changes a 0.5-second repetition interval into 500 without changing the unit. The files may look friendlier but are now technically and semantically inconsistent. A controlled process reverses those changes, attaches a bilingual task glossary, maps original condition codes to explanations, and gives users one canonical link to the immutable data. A qualified researcher then reconstructs a subject/session selection and an event contrast using only the translated guide, while the technical reviewer compares checksums, parsing, row counts and joins.

Suppose the translated guide says the participants attended three clinical visits, but the study protocol says there were three grouped recording sessions, with one session spanning two days. The corrected guide uses the study’s actual timing language rather than assuming that session and visit are synonyms. Suppose it also says a questionnaire measured depression when the instrument was only a self-report mood item. That claim is removed and replaced with a description of the recorded item and its scale. Notice that the most consequential corrections are not grammatical. They preserve the epistemic status of the evidence: what was measured, what was inferred, and what a second investigator could reproduce.

Audit Checklist, Acceptance Criteria and Release Matrix

• All published pages identify a canonical dataset release and declared BIDS version
• No executable filename, suffix, standardized key, code, reference, checksum or binary scientific observation is silently translated
• Numbers, units, precision, arrays, timing origins and coordinate frames remain consistent
• Table columns, participant and session links, stimulus paths and derivative sources resolve
• Every human-readable technical statement is backed by the experiment, codebook or versioned specification
• Sensitive free text is handled only under authorized processing and publication rules
• A target-language researcher can repeat representative discovery and analysis tasks
• Automated validators and manual bilingual scientific QA are both passed
• The released translation records reviewers, date, language, change reason and source version
• A rollback plan, correction contact and monitoring route exist for later dataset releases

The acceptance matrix should mark a field as technically protected, linguistically approved, scientifically reviewed, privacy-cleared and release-tested. Not every field needs every review: a stable filename needs byte equality and technical validation; participant-facing informed consent needs specialized governance and exact ethical review; a plain README paragraph may require a domain-linguist pair. The purpose is proportional assurance, not a ceremonial checklist. Block publication when a severity-one issue could change identity, exposure of personal data, safety-related interpretation or a scientific result.

A Four-Week Implementation Programme

Week 1: Scope and evidence

Freeze the dataset, review governance permissions, identify the BIDS release, inventory all translatable and nontranslatable surfaces, and build the initial field ledger. Select high-consequence examples rather than attempting all strings immediately. For each completed task save a short artifact: manifest, terminology decision, test output, approval record or release note. A future team must be able to tell which evidence supported the decision and what would trigger a re-review.

Week 2: Translation and terminology

Draft approved multilingual descriptions, resolve codebook ambiguities, prepare terms for session, run, event, signal, contrast, derivative and participant categories, and preserve source quotations where needed. For each completed task save a short artifact: manifest, terminology decision, test output, approval record or release note. A future team must be able to tell which evidence supported the decision and what would trigger a re-review.

Week 3: Validator and researcher tests

Run file parsing and schema validation, verify joins and path resolution, compare representative numeric metadata, and ask researchers to carry out blind end-to-end tasks using the localized guide. For each completed task save a short artifact: manifest, terminology decision, test output, approval record or release note. A future team must be able to tell which evidence supported the decision and what would trigger a re-review.

Week 4: Governance and publication

Complete privacy and licensing checks, finish accessibility review, prepare known limitations, link every language view to one dataset identity, record changes and establish a release rollback path. For each completed task save a short artifact: manifest, terminology decision, test output, approval record or release note. A future team must be able to tell which evidence supported the decision and what would trigger a re-review.

Frequently Asked Questions

Can I translate filenames in a BIDS dataset?

Usually no: filenames and defined entity tokens are operational identifiers. Present translated display labels or guides while preserving the canonical file paths. The deciding criterion is whether a reader of the localized text can act on, interpret and trace the same evidence as a reader of the original.

Is every JSON string safe to translate?

No. Some strings are codes, names, file references, versions or constrained values. Read the field definition and the dataset’s declared BIDS version before editing. The deciding criterion is whether a reader of the localized text can act on, interpret and trace the same evidence as a reader of the original.

Can I use machine translation?

For approved non-sensitive explanatory prose it may assist drafting, but review, data-processing permissions, field protection and domain QA remain necessary. The deciding criterion is whether a reader of the localized text can act on, interpret and trace the same evidence as a reader of the original.

Does passing the BIDS Validator mean the translation is correct?

No. Validation targets compliance and structure, not all scientific claims, consent conditions, translation equivalence or human interpretation. The deciding criterion is whether a reader of the localized text can act on, interpret and trace the same evidence as a reader of the original.

Do I convert seconds into milliseconds for target-language readers?

Only in explicitly labeled display text with exact conversion; do not silently alter the numeric machine-readable data. The deciding criterion is whether a reader of the localized text can act on, interpret and trace the same evidence as a reader of the original.

Can a study session be translated as a visit?

Only when the specific protocol supports that equivalence; a BIDS session is a logical grouping and may not match one appointment. The deciding criterion is whether a reader of the localized text can act on, interpret and trace the same evidence as a reader of the original.

Should I translate participant group codes?

Keep the codes as defined. An approved human-readable level description may be localized so long as its mapping remains exact. The deciding criterion is whether a reader of the localized text can act on, interpret and trace the same evidence as a reader of the original.

What is the difference between source data and derivatives?

Source data precede conversion or harmonization, while derivatives result from processing. The original release structure and provenance must make those distinctions intelligible. The deciding criterion is whether a reader of the localized text can act on, interpret and trace the same evidence as a reader of the original.

How do I localize questionnaire responses?

Identify whether the material is a validated instrument, a codebook, a free-text response or a display label; use appropriate rights, adaptation methods and reviews. The deciding criterion is whether a reader of the localized text can act on, interpret and trace the same evidence as a reader of the original.

How should multilingual dashboards handle accessibility?

Use proper language tagging, nontruncated labels, readable tables, stable copied identifiers and tests with screen readers and zoom. The deciding criterion is whether a reader of the localized text can act on, interpret and trace the same evidence as a reader of the original.

What happens when the source study changes?

Treat the new dataset release as a new review event. Diff source content, update affected language views, rerun regression tests and retain prior provenance. The deciding criterion is whether a reader of the localized text can act on, interpret and trace the same evidence as a reader of the original.

What should I do with ambiguous scientific terms?

Keep the uncertainty visible and ask a domain owner; never silently make a measurement more precise or a claim more definite. The deciding criterion is whether a reader of the localized text can act on, interpret and trace the same evidence as a reader of the original.

Reference Sources and Version Notes

The official Brain Imaging Data Structure specification 1.11.2 was consulted on 1 November 2026. The canonical BIDS website explains the standard, its supported modalities, validation ecosystem and extension process. The article references the Common Principles, Dataset Description, Data Summary Files, Phenotypic and Assessment Data, Events and Derivatives Metadata. Statements about a hypothetical study are explanatory examples and are not claims about actual participants or archived datasets.

Standards statements have version boundaries. The stable site was labeled 1.11.2 on the research date; a later update could change a modality’s field rules. A current researcher should consult the release required by the dataset and a version-compatible validator, not adopt a static web paragraph as a substitute for the original specification, institutional requirements or protocol. RFC-style requirement words such as MUST, SHOULD and MAY carry normative distinctions in BIDS documentation. Keep them distinct in translation.

Transfer: From Neuroimaging to Other Research Systems

The method transfers to any field where human language sits on top of machine-readable evidence. Genomic variant identifiers, sensor telemetry, ecological observation codes, museum object records and API specifications all benefit from separating immutable references from translated explanations. The exact data model differs, so reuse the diagnostic questions rather than copying BIDS-specific field rules. Ask what is a key, what is a value, who defines its meaning, which transformation was used, what a reader will infer, and which proof shows that the same fact survived.

A second transfer is organizational. The writer, translator, engineer, scientist and data steward should see one field ledger rather than five contradictory glossaries. The release gate should be proportionate: no participant free text goes to an unapproved processor, no controlled code becomes an untranslated guess, and no scientific conclusion becomes stronger through localization. If a team cannot establish the meaning of a field, a documented unresolved issue is more honest and more useful than a polished sentence that claims unwarranted certainty.

Next Routes in the eduKateSG Translation Architecture

For the general system of meaning, equivalence, contextual judgment, terminology management and human review, return to Master Art of Translation | The Complete System for Moving Meaning Between Languages. This BIDS child owns only the task of cross-language neuroimaging dataset metadata, event and provenance preservation. It does not replace the narrower published rapid BIDS sidecar guide or specialist NIfTI and DICOM resources.

When the challenge is learning words and concept families across subjects, follow the protected Vocabulary Learning Hub. When the challenge is how English sentences use qualifications, reference, modality and evidence, follow the protected How English Works | The Complete Map and Learning Routes. Those hubs retain their existing ownership; they are linked here, not rewritten.

The central rule is simple to remember and demanding to apply: a translated description should lead an international reader to the same participant, the same experiment, the same measurement, the same code and the same supported inference. A neuroimaging dataset is not successfully translated because every visible English word has been replaced. It is successfully translated when the reader’s language changes while the research evidence, its limitations and its reproducibility do not.