VIEW THIS AS

Auto mode follows the Route Engine until you choose a viewpoint.

YOU ARE HERE

ROUTE CHECK

CONNECTED TO

WHAT NEXT

Use the canonical route for this room, or HELP if you are unsure.

Master Art of Translation | The Screening Data Dictionary, Code-System Identity, Value-Set Version and Multilingual Label Fidelity System — How Define, Map, Version and Verify Preserve Clinical Data Meaning

How do you translate a medical screening data dictionary, screening code list, FHIR value set, or multilingual patient label without changing clinical meaning? This world-facing translation guide explains how to preserve the distinction between a stable coded concept, a human-readable display, the version of a controlled vocabulary, and the action a reader is authorised to take. It is written for translators, terminology specialists, clinical information teams, health-content editors, software implementers and reviewers who must localise screening information safely without turning a change of words into a change of medical state.

The hidden problem is that a fluent translation can describe the wrong event. A database field labelled invited becomes booked; a screening-positive signal becomes a confirmed diagnosis; an absent observation becomes a negative test; or a label translated for a new language is accidentally stored as a machine identifier. These errors may pass proofreading while changing eligibility, consent, referral and reporting. The task is therefore semantic interoperability as well as translation: preserve what the source record asserts, what it does not assert, who controls its meaning, and which version makes the interpretation valid.

The repeatable method is to identify the source-of-truth field, define its permitted states and code system, separate machine identity from the public-facing label, lock the effective terminology version, translate under a documented brief, validate boundary cases and then check the actual deployed language. In practice, searches such as translate clinical data dictionary, FHIR terminology localisation, medical screening code translation, SNOMED multilingual preferred terms, and value set version control lead to one underlying question: does the receiving person or system infer exactly the same warranted state from the translated text?

  • Protect the coded meaning (system, code, version, scope), not only wording.
  • Protect epistemic state: unknown, pending, missing, invalid, negative and confirmed are not interchangeable.
  • Protect workflow state: invited, sent, received, accepted, reviewed, authorised, completed and released mean different things.
  • Protect measurement boundaries: units, comparisons, denominators, populations and dates must remain testable.
  • Protect release identity: approval, effective date, fallback, rollback and deployed-language verification require separate evidence.

Architecture: EDKSG-TRANS-MASTER-WORLD-3300. This specialist child addresses data-dictionary and code-system fidelity for multilingual screening information; it does not replace clinical interpretation, a general vocabulary course, or the master translation method. All sample records, terms, values, events and workflows below are deliberately fictional teaching examples; they are not medical instructions, actual code-set mappings or individual clinical recommendations.

1. Search intent and the decision that the translation must preserve

People searching for medical terminology translation sometimes expect a bilingual glossary. A glossary helps only after the writer knows whether the item is a concept, a display string, a field name, a value-set membership rule, a measurement unit, or an instruction. Those are different kinds of things. A glossary entry for positive does not settle whether positive refers to a screening interpretation, a confirmatory finding, an antibody, a quality-control state, or a patient survey answer.

A source data dictionary should declare the field’s purpose, value domain, null handling, provenance, controlled vocabulary, unit, effective date and intended consumers. For multilingual releases it also needs a description of which words are authorised user-facing labels and which strings are protocol or machine keys that must never be translated. The translator should receive these distinctions before drafting rather than discovering them after an incident.

An operational translation brief answers a sequence of questions: What decision does a reader make? What data does the downstream system consume? Which status transitions are valid? Which facts come from a person, a clinician, a device or an algorithm? What confidence is warranted? Which label changes are cosmetic, and which changes alter clinical content? The answers establish the translation’s scope before anyone selects a phrase.

A translated label is not automatically a clinical instruction. It may be an interface prompt, a narrative explanation, a category name, a statistical qualifier or a status of record movement. The target text must communicate its function. A display for experts may be shorter than a public explanation, but both must refer to the same controlled concept and the same allowed action.

The specialist’s first output is often not a translation. It is an inventory of distinctions at risk: source and target audiences, controlled codes, terms that have multiple senses, missing/unknown states, inclusion/exclusion operators, dates, measurements, versions and handoff responsibility. Translation begins once each risk has a designated owner and a test.

2. Hidden problem: four layers that a fluent sentence can collapse

  • Meaning layer: what the field asserts about the world, the participant or the workflow.
  • Representation layer: identifiers, codes, value sets, units, dates, locale tags and optional narrative text.
  • Expression layer: display labels, patient explanations, clinician-facing terms, grammatical scope and register.
  • Governance layer: who approved the terminology, what changed, when it became effective and whether deployment reached users.

The most costly mistakes arise when two layers are treated as substitutes. Translating an identifier can change data identity; keeping an identifier intact but mistranslating its display can mislead a reader. Updating a glossary without synchronising a deployed selection list produces diverging truths. Replacing unknown with negative makes an unsupported assertion. A quality process should therefore test each layer separately and test the relationships between them.

Clinical concepts sometimes have multiple acceptable expressions. SNOMED CT documentation distinguishes a concept from its descriptions and uses language reference sets to identify preferred or acceptable terms by language or dialect. Preferred wording is therefore contextual, rather than a single English phrase that all languages must imitate. An approved synonym should not be mistaken for a separate diagnosis, and an unfamiliar synonym should not silently become a new code.

HL7 FHIR distinguishes a Coding’s system and code from display text and permits CodeableConcept to carry coding alongside human-language text. This distinction is decisive for translation review. The identifier is what participating software should use for coding identity; a person reads a display or explanatory phrase. The system may legitimately present a different display in another language without changing the coded identity. Conversely, when the underlying code changes, changing only a label is inadequate.

A value set is not just a static list copied into a spreadsheet. Its canonical reference, version, expansion parameters and terminology-system versions can matter. A translated dropdown must preserve the approved membership and the relations among concepts. The choice of target-language wording cannot be allowed to silently add an excluded option, omit a valid selection or turn a narrower concept into a broader one.

The question is not whether every interface can display every official term. It is whether the application can document why a particular label was chosen for a particular audience, map it back to an authoritative concept, and recognise when its source or the scope of a value set has changed.

3. Mechanism: map a sentence into a controlled meaning record

Entity. What object or person does the field describe? A specimen ID must not become a participant ID. A reviewer should annotate this dimension explicitly in the source record, check the target-language candidate against the same dimension, and refuse release where the dimension lacks an owner or evidence. The final decision can then be reconstructed even when the surface wording changes across languages.

Predicate. What event or relation is asserted? A referral was sent, not necessarily accepted. A reviewer should annotate this dimension explicitly in the source record, check the target-language candidate against the same dimension, and refuse release where the dimension lacks an owner or evidence. The final decision can then be reconstructed even when the surface wording changes across languages.

Polarity. Is it positive, negative, absent or unknown? No entry is not a negative test. A reviewer should annotate this dimension explicitly in the source record, check the target-language candidate against the same dimension, and refuse release where the dimension lacks an owner or evidence. The final decision can then be reconstructed even when the surface wording changes across languages.

Modality. What is required, possible, permitted or only suggested? May be eligible is not is eligible. A reviewer should annotate this dimension explicitly in the source record, check the target-language candidate against the same dimension, and refuse release where the dimension lacks an owner or evidence. The final decision can then be reconstructed even when the surface wording changes across languages.

Time. Which date anchors the statement and its authority? Effective later is not active now. A reviewer should annotate this dimension explicitly in the source record, check the target-language candidate against the same dimension, and refuse release where the dimension lacks an owner or evidence. The final decision can then be reconstructed even when the surface wording changes across languages.

Scope. Which cohort, method, language and purpose qualify it? A study estimate is not universal. A reviewer should annotate this dimension explicitly in the source record, check the target-language candidate against the same dimension, and refuse release where the dimension lacks an owner or evidence. The final decision can then be reconstructed even when the surface wording changes across languages.

Provenance. Who observed, entered, approved or received it? Patient report is not clinician observation. A reviewer should annotate this dimension explicitly in the source record, check the target-language candidate against the same dimension, and refuse release where the dimension lacks an owner or evidence. The final decision can then be reconstructed even when the surface wording changes across languages.

Version. Which terminology and rule release is operative? A newer glossary does not rewrite historical records. A reviewer should annotate this dimension explicitly in the source record, check the target-language candidate against the same dimension, and refuse release where the dimension lacks an owner or evidence. The final decision can then be reconstructed even when the surface wording changes across languages.

Action. Which next step is valid and who owns it? Being invited is not being booked. A reviewer should annotate this dimension explicitly in the source record, check the target-language candidate against the same dimension, and refuse release where the dimension lacks an owner or evidence. The final decision can then be reconstructed even when the surface wording changes across languages.

Evidence. Which test proves faithful transfer? Reproduce the edge case, not just the sentence. A reviewer should annotate this dimension explicitly in the source record, check the target-language candidate against the same dimension, and refuse release where the dimension lacks an owner or evidence. The final decision can then be reconstructed even when the surface wording changes across languages.

4. Repeatable method: fourteen controlled translation steps

1. Intake and intent

Record the audience, action, data controller, languages, delivery channels and source artefact. Determine whether the item is informational, data-bearing, safety-sensitive or instructional. Deliverable: a brief identifying both the human decision and the downstream machine dependency. The gate is not complete merely because someone marked a task done. A second person should be able to identify the source snapshot, the target asset, the exact acceptance criterion, the evidence of review and the person authorised to release it. If any one element is missing, label the work pending rather than silently advancing its workflow state.

2. Source freeze

Assign an immutable document or data-dictionary revision. Record the field IDs, source labels, code-system versions and application build under review. Deliverable: a reproducible snapshot, not an unversioned email attachment. The gate is not complete merely because someone marked a task done. A second person should be able to identify the source snapshot, the target asset, the exact acceptance criterion, the evidence of review and the person authorised to release it. If any one element is missing, label the work pending rather than silently advancing its workflow state.

3. Field classification

Mark identifiers, coded values, free text, units, dates, numbers, message templates and user interface labels separately. Deliverable: an explicit translate/do-not-translate matrix. The gate is not complete merely because someone marked a task done. A second person should be able to identify the source snapshot, the target asset, the exact acceptance criterion, the evidence of review and the person authorised to release it. If any one element is missing, label the work pending rather than silently advancing its workflow state.

4. Concept inventory

List every sense distinction, negation, condition, scope and action boundary. Ask subject experts where the dictionary is underspecified. Deliverable: a concept register that records unresolved ambiguity instead of guessing. The gate is not complete merely because someone marked a task done. A second person should be able to identify the source snapshot, the target asset, the exact acceptance criterion, the evidence of review and the person authorised to release it. If any one element is missing, label the work pending rather than silently advancing its workflow state.

5. Code-system verification

Bind machine codes to the named system and, where relevant, its version. Do not infer identical meanings from visually similar identifiers. Deliverable: lookup evidence against an authorised terminology source. The gate is not complete merely because someone marked a task done. A second person should be able to identify the source snapshot, the target asset, the exact acceptance criterion, the evidence of review and the person authorised to release it. If any one element is missing, label the work pending rather than silently advancing its workflow state.

6. Value-set constraint

Fix the allowed selection universe, its inclusion/exclusion rules, effective version and expansion assumptions. Deliverable: a version-locked list for multilingual testing. The gate is not complete merely because someone marked a task done. A second person should be able to identify the source snapshot, the target asset, the exact acceptance criterion, the evidence of review and the person authorised to release it. If any one element is missing, label the work pending rather than silently advancing its workflow state.

7. Terminology brief

Give translators preferred terms, acceptable synonyms, forbidden confusions, sample sentences, audiences and register requirements. Deliverable: a controlled brief with explicit concept identifiers. The gate is not complete merely because someone marked a task done. A second person should be able to identify the source snapshot, the target asset, the exact acceptance criterion, the evidence of review and the person authorised to release it. If any one element is missing, label the work pending rather than silently advancing its workflow state.

8. Translation and revision

Produce a target candidate that preserves action, state, time, numbers, uncertainty and provenance. Have an appropriately competent reviewer check meaning independently. Deliverable: tracked changes and resolved terminology queries. The gate is not complete merely because someone marked a task done. A second person should be able to identify the source snapshot, the target asset, the exact acceptance criterion, the evidence of review and the person authorised to release it. If any one element is missing, label the work pending rather than silently advancing its workflow state.

9. Computational validation

Check structural fields, identifiers, units, date formats, placeholders, interpolation boundaries and code/display separation with automated tests. Deliverable: machine-readable pass/fail records. The gate is not complete merely because someone marked a task done. A second person should be able to identify the source snapshot, the target asset, the exact acceptance criterion, the evidence of review and the person authorised to release it. If any one element is missing, label the work pending rather than silently advancing its workflow state.

10. Human scenario validation

Have reviewers enact plausible user decisions from the source and target separately, then compare the warranted conclusions. Deliverable: a decision-equivalence test, not only a word-level score. The gate is not complete merely because someone marked a task done. A second person should be able to identify the source snapshot, the target asset, the exact acceptance criterion, the evidence of review and the person authorised to release it. If any one element is missing, label the work pending rather than silently advancing its workflow state.

11. Boundary testing

Test equality at thresholds, missing vs negative, incomplete vs complete, pending vs approved and old vs new releases. Deliverable: edge-case cases that would catch wrong categorisation. The gate is not complete merely because someone marked a task done. A second person should be able to identify the source snapshot, the target asset, the exact acceptance criterion, the evidence of review and the person authorised to release it. If any one element is missing, label the work pending rather than silently advancing its workflow state.

12. Controlled release

Approve the exact language asset and connect it to the correct build and terminology version. Keep staged, scheduled and deployed states separate. Deliverable: release receipt per channel and target language. The gate is not complete merely because someone marked a task done. A second person should be able to identify the source snapshot, the target asset, the exact acceptance criterion, the evidence of review and the person authorised to release it. If any one element is missing, label the work pending rather than silently advancing its workflow state.

13. Post-release surveillance

Monitor misunderstood labels, terminological drift, clinician feedback and mapping errors without treating every stylistic complaint as a clinical incident. Deliverable: severity-ranked feedback with evidence and a responsible reviewer. The gate is not complete merely because someone marked a task done. A second person should be able to identify the source snapshot, the target asset, the exact acceptance criterion, the evidence of review and the person authorised to release it. If any one element is missing, label the work pending rather than silently advancing its workflow state.

14. Rollback and learning

Retire defective display assets, maintain historical records, record corrections and add new regression tests before reuse. Deliverable: traceable corrective action and proof that affected audiences now see the correct version. The gate is not complete merely because someone marked a task done. A second person should be able to identify the source snapshot, the target asset, the exact acceptance criterion, the evidence of review and the person authorised to release it. If any one element is missing, label the work pending rather than silently advancing its workflow state.

5. Failure modes: what to look for before publishing labels

Code-as-label: a targeted diagnostic

Failure signature: The visible translation is inserted into a machine code field. Why it matters is that the wrong target text may cause a different human decision or machine interpretation even when grammar looks impeccable. The correction is operational rather than cosmetic: Keep the code-system URI and code immutable; localise display only. A translator should document the source assertion, the candidate assertion, the difference in consequence, and the authority approving the repair.

Verification exercise: Switch locale and verify that stored identity remains identical. Use a counterexample in which the tempting wording would create an unsupported state; inspect the output rather than merely checking a box. If the system passes only in one language, the change is not yet resolved. Retest with the old release, a current release, a null field and a realistic downstream recipient to expose hidden dependencies.

Identifier collision: a targeted diagnostic

Failure signature: Two systems reuse the same code string. Why it matters is that the wrong target text may cause a different human decision or machine interpretation even when grammar looks impeccable. The correction is operational rather than cosmetic: Require system namespace and version. A translator should document the source assertion, the candidate assertion, the difference in consequence, and the authority approving the repair.

Verification exercise: Round-trip two identical strings from different systems without merging. Use a counterexample in which the tempting wording would create an unsupported state; inspect the output rather than merely checking a box. If the system passes only in one language, the change is not yet resolved. Retest with the old release, a current release, a null field and a realistic downstream recipient to expose hidden dependencies.

Dictionary drift: a targeted diagnostic

Failure signature: A later source revision is silently mixed with an earlier target. Why it matters is that the wrong target text may cause a different human decision or machine interpretation even when grammar looks impeccable. The correction is operational rather than cosmetic: Lock both to a documented source release. A translator should document the source assertion, the candidate assertion, the difference in consequence, and the authority approving the repair.

Verification exercise: Compare field hashes and effective dates. Use a counterexample in which the tempting wording would create an unsupported state; inspect the output rather than merely checking a box. If the system passes only in one language, the change is not yet resolved. Retest with the old release, a current release, a null field and a realistic downstream recipient to expose hidden dependencies.

False equivalence: a targeted diagnostic

Failure signature: A target term covers a broader category than the source. Why it matters is that the wrong target text may cause a different human decision or machine interpretation even when grammar looks impeccable. The correction is operational rather than cosmetic: Mark mapping relation and obtain domain review. A translator should document the source assertion, the candidate assertion, the difference in consequence, and the authority approving the repair.

Verification exercise: Find a counterexample admitted only by the broader term. Use a counterexample in which the tempting wording would create an unsupported state; inspect the output rather than merely checking a box. If the system passes only in one language, the change is not yet resolved. Retest with the old release, a current release, a null field and a realistic downstream recipient to expose hidden dependencies.

Null collapse: a targeted diagnostic

Failure signature: Unknown, not asked and missing all render as no. Why it matters is that the wrong target text may cause a different human decision or machine interpretation even when grammar looks impeccable. The correction is operational rather than cosmetic: Define separate missingness states. A translator should document the source assertion, the candidate assertion, the difference in consequence, and the authority approving the repair.

Verification exercise: Test source examples for each state. Use a counterexample in which the tempting wording would create an unsupported state; inspect the output rather than merely checking a box. If the system passes only in one language, the change is not yet resolved. Retest with the old release, a current release, a null field and a realistic downstream recipient to expose hidden dependencies.

State leap: a targeted diagnostic

Failure signature: Sent becomes received or invited becomes booked. Why it matters is that the wrong target text may cause a different human decision or machine interpretation even when grammar looks impeccable. The correction is operational rather than cosmetic: Preserve event and evidence of transition. A translator should document the source assertion, the candidate assertion, the difference in consequence, and the authority approving the repair.

Verification exercise: Reconstruct timestamped event history. Use a counterexample in which the tempting wording would create an unsupported state; inspect the output rather than merely checking a box. If the system passes only in one language, the change is not yet resolved. Retest with the old release, a current release, a null field and a realistic downstream recipient to expose hidden dependencies.

Time collapse: a targeted diagnostic

Failure signature: Published becomes effective or scheduled becomes active. Why it matters is that the wrong target text may cause a different human decision or machine interpretation even when grammar looks impeccable. The correction is operational rather than cosmetic: Retain separate temporal fields. A translator should document the source assertion, the candidate assertion, the difference in consequence, and the authority approving the repair.

Verification exercise: Test an encounter occurring between dates. Use a counterexample in which the tempting wording would create an unsupported state; inspect the output rather than merely checking a box. If the system passes only in one language, the change is not yet resolved. Retest with the old release, a current release, a null field and a realistic downstream recipient to expose hidden dependencies.

Threshold operator loss: a targeted diagnostic

Failure signature: Greater than or equal to becomes greater than. Why it matters is that the wrong target text may cause a different human decision or machine interpretation even when grammar looks impeccable. The correction is operational rather than cosmetic: Include mathematical comparison as a protected invariant. A translator should document the source assertion, the candidate assertion, the difference in consequence, and the authority approving the repair.

Verification exercise: Test a measurement exactly on the threshold. Use a counterexample in which the tempting wording would create an unsupported state; inspect the output rather than merely checking a box. If the system passes only in one language, the change is not yet resolved. Retest with the old release, a current release, a null field and a realistic downstream recipient to expose hidden dependencies.

Unit omission: a targeted diagnostic

Failure signature: Displayed quantity loses unit or changes it accidentally. Why it matters is that the wrong target text may cause a different human decision or machine interpretation even when grammar looks impeccable. The correction is operational rather than cosmetic: Keep UCUM or validated equivalent and conversion evidence. A translator should document the source assertion, the candidate assertion, the difference in consequence, and the authority approving the repair.

Verification exercise: Check numeric round-trip for a nontrivial conversion. Use a counterexample in which the tempting wording would create an unsupported state; inspect the output rather than merely checking a box. If the system passes only in one language, the change is not yet resolved. Retest with the old release, a current release, a null field and a realistic downstream recipient to expose hidden dependencies.

Denominator substitution: a targeted diagnostic

Failure signature: Rate among all invitees becomes rate among tested participants. Why it matters is that the wrong target text may cause a different human decision or machine interpretation even when grammar looks impeccable. The correction is operational rather than cosmetic: Preserve numerator and denominator definitions. A translator should document the source assertion, the candidate assertion, the difference in consequence, and the authority approving the repair.

Verification exercise: Recompute metric from controlled sample rows. Use a counterexample in which the tempting wording would create an unsupported state; inspect the output rather than merely checking a box. If the system passes only in one language, the change is not yet resolved. Retest with the old release, a current release, a null field and a realistic downstream recipient to expose hidden dependencies.

Personalisation drift: a targeted diagnostic

Failure signature: Population-level estimate turns into personal certainty. Why it matters is that the wrong target text may cause a different human decision or machine interpretation even when grammar looks impeccable. The correction is operational rather than cosmetic: Preserve study context and conditional direction. A translator should document the source assertion, the candidate assertion, the difference in consequence, and the authority approving the repair.

Verification exercise: Ask whether the candidate claims an individual diagnosis. Use a counterexample in which the tempting wording would create an unsupported state; inspect the output rather than merely checking a box. If the system passes only in one language, the change is not yet resolved. Retest with the old release, a current release, a null field and a realistic downstream recipient to expose hidden dependencies.

Uncertain-to-certain: a targeted diagnostic

Failure signature: Indeterminate becomes positive or negative. Why it matters is that the wrong target text may cause a different human decision or machine interpretation even when grammar looks impeccable. The correction is operational rather than cosmetic: Protect distinct result interpretation codes. A translator should document the source assertion, the candidate assertion, the difference in consequence, and the authority approving the repair.

Verification exercise: Trace the designated next action for each state. Use a counterexample in which the tempting wording would create an unsupported state; inspect the output rather than merely checking a box. If the system passes only in one language, the change is not yet resolved. Retest with the old release, a current release, a null field and a realistic downstream recipient to expose hidden dependencies.

Approval confusion: a targeted diagnostic

Failure signature: Translation approved by linguist looks clinically signed off. Why it matters is that the wrong target text may cause a different human decision or machine interpretation even when grammar looks impeccable. The correction is operational rather than cosmetic: Record linguistic and clinical approvals separately. A translator should document the source assertion, the candidate assertion, the difference in consequence, and the authority approving the repair.

Verification exercise: Inspect who owns each sign-off and its scope. Use a counterexample in which the tempting wording would create an unsupported state; inspect the output rather than merely checking a box. If the system passes only in one language, the change is not yet resolved. Retest with the old release, a current release, a null field and a realistic downstream recipient to expose hidden dependencies.

Fallback disguise: a targeted diagnostic

Failure signature: Unreviewed fallback appears as an approved target-language notice. Why it matters is that the wrong target text may cause a different human decision or machine interpretation even when grammar looks impeccable. The correction is operational rather than cosmetic: Tag served locale and review state. A translator should document the source assertion, the candidate assertion, the difference in consequence, and the authority approving the repair.

Verification exercise: Simulate a missing translation asset. Use a counterexample in which the tempting wording would create an unsupported state; inspect the output rather than merely checking a box. If the system passes only in one language, the change is not yet resolved. Retest with the old release, a current release, a null field and a realistic downstream recipient to expose hidden dependencies.

Language tag absence: a targeted diagnostic

Failure signature: Screen reader or interface does not know content language. Why it matters is that the wrong target text may cause a different human decision or machine interpretation even when grammar looks impeccable. The correction is operational rather than cosmetic: Assign valid language metadata at content and fragment levels. A translator should document the source assertion, the candidate assertion, the difference in consequence, and the authority approving the repair.

Verification exercise: Check accessibility-tree language announcements. Use a counterexample in which the tempting wording would create an unsupported state; inspect the output rather than merely checking a box. If the system passes only in one language, the change is not yet resolved. Retest with the old release, a current release, a null field and a realistic downstream recipient to expose hidden dependencies.

Preferred-term absolutism: a targeted diagnostic

Failure signature: A dialect-specific preferred term becomes mandatory everywhere. Why it matters is that the wrong target text may cause a different human decision or machine interpretation even when grammar looks impeccable. The correction is operational rather than cosmetic: Follow the applicable language reference set. A translator should document the source assertion, the candidate assertion, the difference in consequence, and the authority approving the repair.

Verification exercise: Compare two allowed dialect views of the same concept. Use a counterexample in which the tempting wording would create an unsupported state; inspect the output rather than merely checking a box. If the system passes only in one language, the change is not yet resolved. Retest with the old release, a current release, a null field and a realistic downstream recipient to expose hidden dependencies.

Historical rewrite: a targeted diagnostic

Failure signature: An inactive code disappears from archived reports. Why it matters is that the wrong target text may cause a different human decision or machine interpretation even when grammar looks impeccable. The correction is operational rather than cosmetic: Maintain historical decoding and provenance. A translator should document the source assertion, the candidate assertion, the difference in consequence, and the authority approving the repair.

Verification exercise: Read a prior record with its original version. Use a counterexample in which the tempting wording would create an unsupported state; inspect the output rather than merely checking a box. If the system passes only in one language, the change is not yet resolved. Retest with the old release, a current release, a null field and a realistic downstream recipient to expose hidden dependencies.

Cross-document mismatch: a targeted diagnostic

Failure signature: A consent message and dashboard use different states. Why it matters is that the wrong target text may cause a different human decision or machine interpretation even when grammar looks impeccable. The correction is operational rather than cosmetic: Use one controlled state model and paired verification. A translator should document the source assertion, the candidate assertion, the difference in consequence, and the authority approving the repair.

Verification exercise: Follow one fictional case across every document. Use a counterexample in which the tempting wording would create an unsupported state; inspect the output rather than merely checking a box. If the system passes only in one language, the change is not yet resolved. Retest with the old release, a current release, a null field and a realistic downstream recipient to expose hidden dependencies.

Bad automation: a targeted diagnostic

Failure signature: Machine translation invents an action absent from the original. Why it matters is that the wrong target text may cause a different human decision or machine interpretation even when grammar looks impeccable. The correction is operational rather than cosmetic: Require human review for high-consequence fields. A translator should document the source assertion, the candidate assertion, the difference in consequence, and the authority approving the repair.

Verification exercise: Compare permitted decisions in source and target. Use a counterexample in which the tempting wording would create an unsupported state; inspect the output rather than merely checking a box. If the system passes only in one language, the change is not yet resolved. Retest with the old release, a current release, a null field and a realistic downstream recipient to expose hidden dependencies.

Unverified deployment: a targeted diagnostic

Failure signature: QA passed in staging but public interface still serves old labels. Why it matters is that the wrong target text may cause a different human decision or machine interpretation even when grammar looks impeccable. The correction is operational rather than cosmetic: Require deployment receipts and live checks. A translator should document the source assertion, the candidate assertion, the difference in consequence, and the authority approving the repair.

Verification exercise: Open the production interface in each locale. Use a counterexample in which the tempting wording would create an unsupported state; inspect the output rather than merely checking a box. If the system passes only in one language, the change is not yet resolved. Retest with the old release, a current release, a null field and a realistic downstream recipient to expose hidden dependencies.

Unbounded correction: a targeted diagnostic

Failure signature: One fixed string alters a shared term used in unrelated pathways. Why it matters is that the wrong target text may cause a different human decision or machine interpretation even when grammar looks impeccable. The correction is operational rather than cosmetic: Identify impact radius before changing shared assets. A translator should document the source assertion, the candidate assertion, the difference in consequence, and the authority approving the repair.

Verification exercise: Run all dependent workflow regression suites. Use a counterexample in which the tempting wording would create an unsupported state; inspect the output rather than merely checking a box. If the system passes only in one language, the change is not yet resolved. Retest with the old release, a current release, a null field and a realistic downstream recipient to expose hidden dependencies.

Unowned mapping: a targeted diagnostic

Failure signature: Integrator guesses a code-system crosswalk. Why it matters is that the wrong target text may cause a different human decision or machine interpretation even when grammar looks impeccable. The correction is operational rather than cosmetic: Escalate to accountable terminology authority. A translator should document the source assertion, the candidate assertion, the difference in consequence, and the authority approving the repair.

Verification exercise: Demand an evidence source for each accepted mapping. Use a counterexample in which the tempting wording would create an unsupported state; inspect the output rather than merely checking a box. If the system passes only in one language, the change is not yet resolved. Retest with the old release, a current release, a null field and a realistic downstream recipient to expose hidden dependencies.

Patient quotation erasure: a targeted diagnostic

Failure signature: A person’s qualified answer is forced into a binary value. Why it matters is that the wrong target text may cause a different human decision or machine interpretation even when grammar looks impeccable. The correction is operational rather than cosmetic: Preserve free text where permitted and flag coding uncertainty. A translator should document the source assertion, the candidate assertion, the difference in consequence, and the authority approving the repair.

Verification exercise: Compare original quotation with final coded assertion. Use a counterexample in which the tempting wording would create an unsupported state; inspect the output rather than merely checking a box. If the system passes only in one language, the change is not yet resolved. Retest with the old release, a current release, a null field and a realistic downstream recipient to expose hidden dependencies.

Privacy overclaim: a targeted diagnostic

Failure signature: Pseudonymisation is presented as irreversible anonymisation. Why it matters is that the wrong target text may cause a different human decision or machine interpretation even when grammar looks impeccable. The correction is operational rather than cosmetic: Describe the actual transformation and residual risk. A translator should document the source assertion, the candidate assertion, the difference in consequence, and the authority approving the repair.

Verification exercise: Review linkage keys and indirect identifiers. Use a counterexample in which the tempting wording would create an unsupported state; inspect the output rather than merely checking a box. If the system passes only in one language, the change is not yet resolved. Retest with the old release, a current release, a null field and a realistic downstream recipient to expose hidden dependencies.

Premature closure: a targeted diagnostic

Failure signature: Local workflow is closed though referral remains open elsewhere. Why it matters is that the wrong target text may cause a different human decision or machine interpretation even when grammar looks impeccable. The correction is operational rather than cosmetic: Require external acceptance or completion evidence. A translator should document the source assertion, the candidate assertion, the difference in consequence, and the authority approving the repair.

Verification exercise: Trace final destination acknowledgement. Use a counterexample in which the tempting wording would create an unsupported state; inspect the output rather than merely checking a box. If the system passes only in one language, the change is not yet resolved. Retest with the old release, a current release, a null field and a realistic downstream recipient to expose hidden dependencies.

Lost rollback: a targeted diagnostic

Failure signature: New label cannot be reverted or historic state reconstructed. Why it matters is that the wrong target text may cause a different human decision or machine interpretation even when grammar looks impeccable. The correction is operational rather than cosmetic: Archive approved assets and deployment mapping. A translator should document the source assertion, the candidate assertion, the difference in consequence, and the authority approving the repair.

Verification exercise: Restore the prior release in a rehearsal. Use a counterexample in which the tempting wording would create an unsupported state; inspect the output rather than merely checking a box. If the system passes only in one language, the change is not yet resolved. Retest with the old release, a current release, a null field and a realistic downstream recipient to expose hidden dependencies.

6. Worked diagnostics: fictional screening data-dictionary and localisation cases

Each diagnostic below is a deliberately constructed language-neutral exercise, not a published clinical instruction or an authoritative coding mapping. The source phrase and misleading rendering are in English so that readers working with any language pair can test the semantic difference before deciding how their own target language must express it. Each case follows one field through a public label, a data interchange and a controlled release. This three-view method makes the same semantic error visible in different forms rather than offering a disconnected glossary.

1. Invitation state: three-view fidelity diagnostic

Source record: “invited for screening.” Misleading candidate: “booked for screening.” Controlled interpretation: “invitation sent; appointment not yet booked.” This is not a style preference. An invitation is an outreach event, not evidence that the recipient scheduled a visit. A person may wait at home believing an appointment exists. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

1A. Public-facing translation and human decision

Consider a plain-language explanation on an invitation page, where the recipient has limited context. The source label is “invited for screening”; an apparently natural localised phrase would assert “booked for screening”. The approved meaning to carry is “invitation sent; appointment not yet booked”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, a person may wait at home believing an appointment exists. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Can the scheduler prove a booking exists?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

1B. Machine-readable code and data-mapping test

Move the example from a public screen to an export destined for a reporting system. The relevant mapping inventory is invite_status as event, appointment_status as separate field. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “booked for screening” masquerading as “invited for screening”; the accepted reading remains “invitation sent; appointment not yet booked”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

1C. Version, correction and post-release verification

Now suppose a translator proposes a better term after the interface has gone live. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. An invitation is an outreach event, not evidence that the recipient scheduled a visit. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. A person may wait at home believing an appointment exists. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Can the scheduler prove a booking exists?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for invitation state: Work independently in two roles. The first reviewer is given only “invited for screening” and must write its weakest warranted assertion; the second reviewer is given only “invitation sent; appointment not yet booked” in the target language and must do the same. Compare the two assertions, then introduce the tempting “booked for screening” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

2. Eligibility assessment: three-view fidelity diagnostic

Source record: “eligibility not yet determined.” Misleading candidate: “ineligible.” Controlled interpretation: “eligibility review pending.” This is not a style preference. Unknown eligibility must not be compressed into a negative eligibility decision. A person could incorrectly abandon further assessment. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

2A. Public-facing translation and human decision

Picture a multilingual dashboard whose short label must fit in a compact interface. The source label is “eligibility not yet determined”; an apparently natural localised phrase would assert “ineligible”. The approved meaning to carry is “eligibility review pending”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, a person could incorrectly abandon further assessment. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Is there evidence for the final state?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

2B. Machine-readable code and data-mapping test

Next test the field in a machine-to-machine exchange during reconciliation. The relevant mapping inventory is eligibility_state with pending, eligible and ineligible values. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “ineligible” masquerading as “eligibility not yet determined”; the accepted reading remains “eligibility review pending”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

2C. Version, correction and post-release verification

Consider a complaint claiming that the displayed wording led to a different next action. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Unknown eligibility must not be compressed into a negative eligibility decision. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. A person could incorrectly abandon further assessment. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Is there evidence for the final state?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for eligibility assessment: Work independently in two roles. The first reviewer is given only “eligibility not yet determined” and must write its weakest warranted assertion; the second reviewer is given only “eligibility review pending” in the target language and must do the same. Compare the two assertions, then introduce the tempting “ineligible” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

3. Age criterion: three-view fidelity diagnostic

Source record: “eligible on the reference date.” Misleading candidate: “eligible from birthday onward.” Controlled interpretation: “age assessed using the named reference date.” This is not a style preference. The temporal anchor is part of the rule; a birthday alone does not determine the evaluation day. Two systems might disagree around a date boundary. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

3A. Public-facing translation and human decision

Consider a notification that a non-specialist reads before deciding whether to contact a service. The source label is “eligible on the reference date”; an apparently natural localised phrase would assert “eligible from birthday onward”. The approved meaning to carry is “age assessed using the named reference date”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, two systems might disagree around a date boundary. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Which date did the rule use?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

3B. Machine-readable code and data-mapping test

Consider the same state inside a data warehouse that computes aggregate screening counts. The relevant mapping inventory is reference_date, birth_date and eligibility_rule_version. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “eligible from birthday onward” masquerading as “eligible on the reference date”; the accepted reading remains “age assessed using the named reference date”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

3C. Version, correction and post-release verification

Imagine a controlled rollback while older records remain available. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. The temporal anchor is part of the rule; a birthday alone does not determine the evaluation day. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. Two systems might disagree around a date boundary. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Which date did the rule use?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for age criterion: Work independently in two roles. The first reviewer is given only “eligible on the reference date” and must write its weakest warranted assertion; the second reviewer is given only “age assessed using the named reference date” in the target language and must do the same. Compare the two assertions, then introduce the tempting “eligible from birthday onward” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

4. Participant identifier: three-view fidelity diagnostic

Source record: “screening participant code.” Misleading candidate: “passport or national identity number.” Controlled interpretation: “pseudonymous screening record identifier.” This is not a style preference. A study or programme identifier may be pseudonymous and must not be treated as a legal identity. Unintended identification or failed record linkage can result. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

4A. Public-facing translation and human decision

Imagine a telephone support agent reading a translated workflow message to a participant. The source label is “screening participant code”; an apparently natural localised phrase would assert “passport or national identity number”. The approved meaning to carry is “pseudonymous screening record identifier”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, unintended identification or failed record linkage can result. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Which system issued the identifier?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

4B. Machine-readable code and data-mapping test

Now imagine a terminology service feeding a multilingual selection list. The relevant mapping inventory is identifier.system, identifier.value and assigning authority. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “passport or national identity number” masquerading as “screening participant code”; the accepted reading remains “pseudonymous screening record identifier”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

4C. Version, correction and post-release verification

Suppose two language variants receive the new asset at different moments. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. A study or programme identifier may be pseudonymous and must not be treated as a legal identity. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. Unintended identification or failed record linkage can result. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Which system issued the identifier?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for participant identifier: Work independently in two roles. The first reviewer is given only “screening participant code” and must write its weakest warranted assertion; the second reviewer is given only “pseudonymous screening record identifier” in the target language and must do the same. Compare the two assertions, then introduce the tempting “passport or national identity number” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

5. Invitation dispatch: three-view fidelity diagnostic

Source record: “invitation generated.” Misleading candidate: “invitation delivered.” Controlled interpretation: “invitation produced; delivery not confirmed.” This is not a style preference. Creation, transmission and receipt are different events with different evidence. The programme may count people reached who never received a message. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

5A. Public-facing translation and human decision

Picture a printed summary that must remain intelligible long after the original encounter. The source label is “invitation generated”; an apparently natural localised phrase would assert “invitation delivered”. The approved meaning to carry is “invitation produced; delivery not confirmed”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, the programme may count people reached who never received a message. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Which timestamp proves delivery?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

5B. Machine-readable code and data-mapping test

Next review the value inside a structured report copied across two organisations. The relevant mapping inventory is generated_at, dispatched_at and delivered_at. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “invitation delivered” masquerading as “invitation generated”; the accepted reading remains “invitation produced; delivery not confirmed”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

5C. Version, correction and post-release verification

Consider a reviewer finding that the source data dictionary was underspecified. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Creation, transmission and receipt are different events with different evidence. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. The programme may count people reached who never received a message. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Which timestamp proves delivery?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for invitation dispatch: Work independently in two roles. The first reviewer is given only “invitation generated” and must write its weakest warranted assertion; the second reviewer is given only “invitation produced; delivery not confirmed” in the target language and must do the same. Compare the two assertions, then introduce the tempting “invitation delivered” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

6. Contact preference: three-view fidelity diagnostic

Source record: “preferred written language.” Misleading candidate: “language legally required for all records.” Controlled interpretation: “recipient’s stated written-language preference.” This is not a style preference. A preference for an interface language is not a universal mandate for all clinical records. Readers may get the wrong script or an inaccessible channel. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

6A. Public-facing translation and human decision

Consider a web form presented alongside a next-step button. The source label is “preferred written language”; an apparently natural localised phrase would assert “language legally required for all records”. The approved meaning to carry is “recipient’s stated written-language preference”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, readers may get the wrong script or an inaccessible channel. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Whose preference was recorded and when?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

6B. Machine-readable code and data-mapping test

Trace the record through a batch migration in which some values are absent. The relevant mapping inventory is Communication.language and preference capture source. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “language legally required for all records” masquerading as “preferred written language”; the accepted reading remains “recipient’s stated written-language preference”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

6C. Version, correction and post-release verification

Imagine a fallback string appearing before the new locale is fully approved. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. A preference for an interface language is not a universal mandate for all clinical records. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. Readers may get the wrong script or an inaccessible channel. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Whose preference was recorded and when?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for contact preference: Work independently in two roles. The first reviewer is given only “preferred written language” and must write its weakest warranted assertion; the second reviewer is given only “recipient’s stated written-language preference” in the target language and must do the same. Compare the two assertions, then introduce the tempting “language legally required for all records” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

7. Screening offer: three-view fidelity diagnostic

Source record: “screening available.” Misleading candidate: “screening recommended for this individual.” Controlled interpretation: “an offer of screening subject to applicable eligibility and choice.” This is not a style preference. An availability statement is weaker than personalised clinical advice. An informational campaign may be mistaken for individual medical instruction. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

7A. Public-facing translation and human decision

Imagine a reminder sent through a channel that does not display the full underlying code. The source label is “screening available”; an apparently natural localised phrase would assert “screening recommended for this individual”. The approved meaning to carry is “an offer of screening subject to applicable eligibility and choice”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, an informational campaign may be mistaken for individual medical instruction. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Does the text attribute an individual recommendation that never occurred?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

7B. Machine-readable code and data-mapping test

Follow the field through an audit log and a derived analytics view. The relevant mapping inventory is offer_type and recommendation_status separated. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “screening recommended for this individual” masquerading as “screening available”; the accepted reading remains “an offer of screening subject to applicable eligibility and choice”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

7C. Version, correction and post-release verification

Suppose an integration test passes but a downstream recipient reports a mismatch. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. An availability statement is weaker than personalised clinical advice. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. An informational campaign may be mistaken for individual medical instruction. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Does the text attribute an individual recommendation that never occurred?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for screening offer: Work independently in two roles. The first reviewer is given only “screening available” and must write its weakest warranted assertion; the second reviewer is given only “an offer of screening subject to applicable eligibility and choice” in the target language and must do the same. Compare the two assertions, then introduce the tempting “screening recommended for this individual” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

8. Consent status: three-view fidelity diagnostic

Source record: “consent discussion completed.” Misleading candidate: “consent granted.” Controlled interpretation: “discussion completed; decision not recorded.” This is not a style preference. Holding a conversation does not prove a participant gave informed permission. Data could be used without the actual recorded decision. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

8A. Public-facing translation and human decision

Imagine a patient-facing portal after an automated screening workflow has produced a status message. The source label is “consent discussion completed”; an apparently natural localised phrase would assert “consent granted”. The approved meaning to carry is “discussion completed; decision not recorded”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, data could be used without the actual recorded decision. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Where is the positive evidence of consent?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

8B. Machine-readable code and data-mapping test

Now follow the same record into an interoperability message handled by software rather than a reader. The relevant mapping inventory is consent_event, consent_status and recorded_at. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “consent granted” masquerading as “consent discussion completed”; the accepted reading remains “discussion completed; decision not recorded”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

8C. Version, correction and post-release verification

Finally imagine a new terminology or software release affecting this field. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Holding a conversation does not prove a participant gave informed permission. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. Data could be used without the actual recorded decision. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Where is the positive evidence of consent?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for consent status: Work independently in two roles. The first reviewer is given only “consent discussion completed” and must write its weakest warranted assertion; the second reviewer is given only “discussion completed; decision not recorded” in the target language and must do the same. Compare the two assertions, then introduce the tempting “consent granted” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

9. Consent withdrawal: three-view fidelity diagnostic

Source record: “withdrawal request received.” Misleading candidate: “all data erased immediately.” Controlled interpretation: “withdrawal received; consequences subject to applicable policy.” This is not a style preference. Withdrawal, data retention and downstream legal handling are separate operations. People may receive impossible promises about deletion. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

9A. Public-facing translation and human decision

Consider a plain-language explanation on an invitation page, where the recipient has limited context. The source label is “withdrawal request received”; an apparently natural localised phrase would assert “all data erased immediately”. The approved meaning to carry is “withdrawal received; consequences subject to applicable policy”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, people may receive impossible promises about deletion. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Was withdrawal processed or only submitted?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

9B. Machine-readable code and data-mapping test

Move the example from a public screen to an export destined for a reporting system. The relevant mapping inventory is withdrawal_received_at and processing_state. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “all data erased immediately” masquerading as “withdrawal request received”; the accepted reading remains “withdrawal received; consequences subject to applicable policy”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

9C. Version, correction and post-release verification

Now suppose a translator proposes a better term after the interface has gone live. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Withdrawal, data retention and downstream legal handling are separate operations. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. People may receive impossible promises about deletion. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Was withdrawal processed or only submitted?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for consent withdrawal: Work independently in two roles. The first reviewer is given only “withdrawal request received” and must write its weakest warranted assertion; the second reviewer is given only “withdrawal received; consequences subject to applicable policy” in the target language and must do the same. Compare the two assertions, then introduce the tempting “all data erased immediately” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

10. Specimen requested: three-view fidelity diagnostic

Source record: “specimen collection requested.” Misleading candidate: “specimen received.” Controlled interpretation: “collection requested, not yet completed.” This is not a style preference. A requested action must remain distinct from a specimen in hand. Laboratories may expect nonexistent material. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

10A. Public-facing translation and human decision

Picture a multilingual dashboard whose short label must fit in a compact interface. The source label is “specimen collection requested”; an apparently natural localised phrase would assert “specimen received”. The approved meaning to carry is “collection requested, not yet completed”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, laboratories may expect nonexistent material. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Can a collection receipt be identified?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

10B. Machine-readable code and data-mapping test

Next test the field in a machine-to-machine exchange during reconciliation. The relevant mapping inventory is request_id and specimen_status. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “specimen received” masquerading as “specimen collection requested”; the accepted reading remains “collection requested, not yet completed”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

10C. Version, correction and post-release verification

Consider a complaint claiming that the displayed wording led to a different next action. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. A requested action must remain distinct from a specimen in hand. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. Laboratories may expect nonexistent material. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Can a collection receipt be identified?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for specimen requested: Work independently in two roles. The first reviewer is given only “specimen collection requested” and must write its weakest warranted assertion; the second reviewer is given only “collection requested, not yet completed” in the target language and must do the same. Compare the two assertions, then introduce the tempting “specimen received” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

11. Collection method: three-view fidelity diagnostic

Source record: “self-collected specimen.” Misleading candidate: “clinician-collected specimen.” Controlled interpretation: “self-collected specimen according to the approved instructions.” This is not a style preference. The actor, not just the specimen, is part of the method definition. Instructions may route participants to the wrong procedure. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

11A. Public-facing translation and human decision

Consider a notification that a non-specialist reads before deciding whether to contact a service. The source label is “self-collected specimen”; an apparently natural localised phrase would assert “clinician-collected specimen”. The approved meaning to carry is “self-collected specimen according to the approved instructions”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, instructions may route participants to the wrong procedure. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Who is authorized to collect it?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

11B. Machine-readable code and data-mapping test

Consider the same state inside a data warehouse that computes aggregate screening counts. The relevant mapping inventory is method_code, collector_role and instructions_version. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “clinician-collected specimen” masquerading as “self-collected specimen”; the accepted reading remains “self-collected specimen according to the approved instructions”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

11C. Version, correction and post-release verification

Imagine a controlled rollback while older records remain available. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. The actor, not just the specimen, is part of the method definition. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. Instructions may route participants to the wrong procedure. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Who is authorized to collect it?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for collection method: Work independently in two roles. The first reviewer is given only “self-collected specimen” and must write its weakest warranted assertion; the second reviewer is given only “self-collected specimen according to the approved instructions” in the target language and must do the same. Compare the two assertions, then introduce the tempting “clinician-collected specimen” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

12. Specimen transport: three-view fidelity diagnostic

Source record: “sample in transit.” Misleading candidate: “sample accepted.” Controlled interpretation: “sample sent; acceptance not recorded.” This is not a style preference. Transport and acceptance are separate custody events. A rejected sample may be mistakenly treated as usable. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

12A. Public-facing translation and human decision

Imagine a telephone support agent reading a translated workflow message to a participant. The source label is “sample in transit”; an apparently natural localised phrase would assert “sample accepted”. The approved meaning to carry is “sample sent; acceptance not recorded”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, a rejected sample may be mistakenly treated as usable. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Who confirmed the specimen was accepted?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

12B. Machine-readable code and data-mapping test

Now imagine a terminology service feeding a multilingual selection list. The relevant mapping inventory is transport_state and laboratory_acceptance_event. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “sample accepted” masquerading as “sample in transit”; the accepted reading remains “sample sent; acceptance not recorded”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

12C. Version, correction and post-release verification

Suppose two language variants receive the new asset at different moments. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Transport and acceptance are separate custody events. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. A rejected sample may be mistakenly treated as usable. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Who confirmed the specimen was accepted?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for specimen transport: Work independently in two roles. The first reviewer is given only “sample in transit” and must write its weakest warranted assertion; the second reviewer is given only “sample sent; acceptance not recorded” in the target language and must do the same. Compare the two assertions, then introduce the tempting “sample accepted” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

13. Test purpose: three-view fidelity diagnostic

Source record: “screening test.” Misleading candidate: “definitive diagnostic test.” Controlled interpretation: “screening test that may require diagnostic follow-up.” This is not a style preference. The test’s purpose governs the claims a result can support. A reader may interpret a screening signal as a definitive diagnosis. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

13A. Public-facing translation and human decision

Picture a printed summary that must remain intelligible long after the original encounter. The source label is “screening test”; an apparently natural localised phrase would assert “definitive diagnostic test”. The approved meaning to carry is “screening test that may require diagnostic follow-up”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, a reader may interpret a screening signal as a definitive diagnosis. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Does this field assert diagnosis or detection for further assessment?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

13B. Machine-readable code and data-mapping test

Next review the value inside a structured report copied across two organisations. The relevant mapping inventory is test_role and follow_up_rule. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “definitive diagnostic test” masquerading as “screening test”; the accepted reading remains “screening test that may require diagnostic follow-up”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

13C. Version, correction and post-release verification

Consider a reviewer finding that the source data dictionary was underspecified. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. The test’s purpose governs the claims a result can support. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. A reader may interpret a screening signal as a definitive diagnosis. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Does this field assert diagnosis or detection for further assessment?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for test purpose: Work independently in two roles. The first reviewer is given only “screening test” and must write its weakest warranted assertion; the second reviewer is given only “screening test that may require diagnostic follow-up” in the target language and must do the same. Compare the two assertions, then introduce the tempting “definitive diagnostic test” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

14. Method identity: three-view fidelity diagnostic

Source record: “method revision B.” Misleading candidate: “same method as revision A.” Controlled interpretation: “method revision B, comparability subject to verification.” This is not a style preference. A label shared by methods does not establish analytical equivalence. Trend charts may merge non-comparable observations. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

14A. Public-facing translation and human decision

Consider a web form presented alongside a next-step button. The source label is “method revision B”; an apparently natural localised phrase would assert “same method as revision A”. The approved meaning to carry is “method revision B, comparability subject to verification”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, trend charts may merge non-comparable observations. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Has comparability been verified rather than assumed?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

14B. Machine-readable code and data-mapping test

Trace the record through a batch migration in which some values are absent. The relevant mapping inventory is method_id, revision and comparability_status. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “same method as revision A” masquerading as “method revision B”; the accepted reading remains “method revision B, comparability subject to verification”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

14C. Version, correction and post-release verification

Imagine a fallback string appearing before the new locale is fully approved. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. A label shared by methods does not establish analytical equivalence. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. Trend charts may merge non-comparable observations. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Has comparability been verified rather than assumed?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for method identity: Work independently in two roles. The first reviewer is given only “method revision B” and must write its weakest warranted assertion; the second reviewer is given only “method revision B, comparability subject to verification” in the target language and must do the same. Compare the two assertions, then introduce the tempting “same method as revision A” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

15. Threshold boundary: three-view fidelity diagnostic

Source record: “result at or above the threshold.” Misleading candidate: “result above the threshold only.” Controlled interpretation: “result greater than or equal to the declared threshold.” This is not a style preference. The inclusion of equality changes classification at exactly the boundary. A single score at the limit may be misclassified. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

15A. Public-facing translation and human decision

Imagine a reminder sent through a channel that does not display the full underlying code. The source label is “result at or above the threshold”; an apparently natural localised phrase would assert “result above the threshold only”. The approved meaning to carry is “result greater than or equal to the declared threshold”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, a single score at the limit may be misclassified. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Does a test case exactly at the boundary pass?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

15B. Machine-readable code and data-mapping test

Follow the field through an audit log and a derived analytics view. The relevant mapping inventory is operator, threshold_value, unit and version. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “result above the threshold only” masquerading as “result at or above the threshold”; the accepted reading remains “result greater than or equal to the declared threshold”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

15C. Version, correction and post-release verification

Suppose an integration test passes but a downstream recipient reports a mismatch. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. The inclusion of equality changes classification at exactly the boundary. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. A single score at the limit may be misclassified. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Does a test case exactly at the boundary pass?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for threshold boundary: Work independently in two roles. The first reviewer is given only “result at or above the threshold” and must write its weakest warranted assertion; the second reviewer is given only “result greater than or equal to the declared threshold” in the target language and must do the same. Compare the two assertions, then introduce the tempting “result above the threshold only” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

16. Unit representation: three-view fidelity diagnostic

Source record: “result 0.8 units per millilitre.” Misleading candidate: “result 0.8 units per litre.” Controlled interpretation: “result 0.8 units per millilitre; keep source unit intact.” This is not a style preference. Numbers and their units form one measurement; rewriting units without conversion changes magnitude. The apparent result may change a classification. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

16A. Public-facing translation and human decision

Imagine a patient-facing portal after an automated screening workflow has produced a status message. The source label is “result 0.8 units per millilitre”; an apparently natural localised phrase would assert “result 0.8 units per litre”. The approved meaning to carry is “result 0.8 units per millilitre; keep source unit intact”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, the apparent result may change a classification. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Was conversion explicitly performed and validated?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

16B. Machine-readable code and data-mapping test

Now follow the same record into an interoperability message handled by software rather than a reader. The relevant mapping inventory is value, UCUM unit and source-unit provenance. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “result 0.8 units per litre” masquerading as “result 0.8 units per millilitre”; the accepted reading remains “result 0.8 units per millilitre; keep source unit intact”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

16C. Version, correction and post-release verification

Finally imagine a new terminology or software release affecting this field. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Numbers and their units form one measurement; rewriting units without conversion changes magnitude. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. The apparent result may change a classification. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Was conversion explicitly performed and validated?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for unit representation: Work independently in two roles. The first reviewer is given only “result 0.8 units per millilitre” and must write its weakest warranted assertion; the second reviewer is given only “result 0.8 units per millilitre; keep source unit intact” in the target language and must do the same. Compare the two assertions, then introduce the tempting “result 0.8 units per litre” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

17. Reference interval: three-view fidelity diagnostic

Source record: “reference interval for a specified method.” Misleading candidate: “universal normal range.” Controlled interpretation: “method-specific reference interval where supplied.” This is not a style preference. The interval may depend on method and population; universalising its scope is unjustified. Readers may compare incompatible ranges. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

17A. Public-facing translation and human decision

Consider a plain-language explanation on an invitation page, where the recipient has limited context. The source label is “reference interval for a specified method”; an apparently natural localised phrase would assert “universal normal range”. The approved meaning to carry is “method-specific reference interval where supplied”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, readers may compare incompatible ranges. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “What makes this interval applicable?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

17B. Machine-readable code and data-mapping test

Move the example from a public screen to an export destined for a reporting system. The relevant mapping inventory is referenceRange, method and scope. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “universal normal range” masquerading as “reference interval for a specified method”; the accepted reading remains “method-specific reference interval where supplied”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

17C. Version, correction and post-release verification

Now suppose a translator proposes a better term after the interface has gone live. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. The interval may depend on method and population; universalising its scope is unjustified. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. Readers may compare incompatible ranges. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “What makes this interval applicable?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for reference interval: Work independently in two roles. The first reviewer is given only “reference interval for a specified method” and must write its weakest warranted assertion; the second reviewer is given only “method-specific reference interval where supplied” in the target language and must do the same. Compare the two assertions, then introduce the tempting “universal normal range” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

18. Missing result: three-view fidelity diagnostic

Source record: “no result recorded.” Misleading candidate: “negative result.” Controlled interpretation: “no result available; result status unresolved.” This is not a style preference. Absence of a value is not a negative measurement. A person may wrongly conclude there is no concerning signal. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

18A. Public-facing translation and human decision

Picture a multilingual dashboard whose short label must fit in a compact interface. The source label is “no result recorded”; an apparently natural localised phrase would assert “negative result”. The approved meaning to carry is “no result available; result status unresolved”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, a person may wrongly conclude there is no concerning signal. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “What evidence distinguishes not measured from negative?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

18B. Machine-readable code and data-mapping test

Next test the field in a machine-to-machine exchange during reconciliation. The relevant mapping inventory is dataAbsentReason versus coded test outcome. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “negative result” masquerading as “no result recorded”; the accepted reading remains “no result available; result status unresolved”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

18C. Version, correction and post-release verification

Consider a complaint claiming that the displayed wording led to a different next action. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Absence of a value is not a negative measurement. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. A person may wrongly conclude there is no concerning signal. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “What evidence distinguishes not measured from negative?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for missing result: Work independently in two roles. The first reviewer is given only “no result recorded” and must write its weakest warranted assertion; the second reviewer is given only “no result available; result status unresolved” in the target language and must do the same. Compare the two assertions, then introduce the tempting “negative result” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

19. Invalid result: three-view fidelity diagnostic

Source record: “test invalid.” Misleading candidate: “test negative.” Controlled interpretation: “invalid result; repeat or follow-up as directed by the programme.” This is not a style preference. An invalid attempt cannot be converted into a valid negative outcome. A repeat pathway may be missed. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

19A. Public-facing translation and human decision

Consider a notification that a non-specialist reads before deciding whether to contact a service. The source label is “test invalid”; an apparently natural localised phrase would assert “test negative”. The approved meaning to carry is “invalid result; repeat or follow-up as directed by the programme”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, a repeat pathway may be missed. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “What action is required for invalidity?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

19B. Machine-readable code and data-mapping test

Consider the same state inside a data warehouse that computes aggregate screening counts. The relevant mapping inventory is result_status and invalid_reason. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “test negative” masquerading as “test invalid”; the accepted reading remains “invalid result; repeat or follow-up as directed by the programme”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

19C. Version, correction and post-release verification

Imagine a controlled rollback while older records remain available. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. An invalid attempt cannot be converted into a valid negative outcome. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. A repeat pathway may be missed. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “What action is required for invalidity?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for invalid result: Work independently in two roles. The first reviewer is given only “test invalid” and must write its weakest warranted assertion; the second reviewer is given only “invalid result; repeat or follow-up as directed by the programme” in the target language and must do the same. Compare the two assertions, then introduce the tempting “test negative” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

20. Indeterminate category: three-view fidelity diagnostic

Source record: “indeterminate.” Misleading candidate: “positive.” Controlled interpretation: “indeterminate; a conclusive category has not been assigned.” This is not a style preference. Uncertainty is a defined reporting state rather than an editor’s weakness. The next procedural step may be skipped. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

20A. Public-facing translation and human decision

Imagine a telephone support agent reading a translated workflow message to a participant. The source label is “indeterminate”; an apparently natural localised phrase would assert “positive”. The approved meaning to carry is “indeterminate; a conclusive category has not been assigned”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, the next procedural step may be skipped. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Does the codebook permit this intermediate state?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

20B. Machine-readable code and data-mapping test

Now imagine a terminology service feeding a multilingual selection list. The relevant mapping inventory is interpretation_code and uncertainty_definition. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “positive” masquerading as “indeterminate”; the accepted reading remains “indeterminate; a conclusive category has not been assigned”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

20C. Version, correction and post-release verification

Suppose two language variants receive the new asset at different moments. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Uncertainty is a defined reporting state rather than an editor’s weakness. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. The next procedural step may be skipped. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Does the codebook permit this intermediate state?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for indeterminate category: Work independently in two roles. The first reviewer is given only “indeterminate” and must write its weakest warranted assertion; the second reviewer is given only “indeterminate; a conclusive category has not been assigned” in the target language and must do the same. Compare the two assertions, then introduce the tempting “positive” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

21. Borderline band: three-view fidelity diagnostic

Source record: “borderline under current method.” Misleading candidate: “all results in this band are positive.” Controlled interpretation: “borderline under the named decision rule.” This is not a style preference. A borderline band may be a separate classification with its own next action. A binary label could trigger incorrect routing. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

21A. Public-facing translation and human decision

Picture a printed summary that must remain intelligible long after the original encounter. The source label is “borderline under current method”; an apparently natural localised phrase would assert “all results in this band are positive”. The approved meaning to carry is “borderline under the named decision rule”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, a binary label could trigger incorrect routing. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Is the intermediate category preserved?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

21B. Machine-readable code and data-mapping test

Next review the value inside a structured report copied across two organisations. The relevant mapping inventory is decision_band, rule_version and next_action. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “all results in this band are positive” masquerading as “borderline under current method”; the accepted reading remains “borderline under the named decision rule”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

21C. Version, correction and post-release verification

Consider a reviewer finding that the source data dictionary was underspecified. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. A borderline band may be a separate classification with its own next action. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. A binary label could trigger incorrect routing. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Is the intermediate category preserved?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for borderline band: Work independently in two roles. The first reviewer is given only “borderline under current method” and must write its weakest warranted assertion; the second reviewer is given only “borderline under the named decision rule” in the target language and must do the same. Compare the two assertions, then introduce the tempting “all results in this band are positive” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

22. Screen-positive signal: three-view fidelity diagnostic

Source record: “screen positive.” Misleading candidate: “diagnosed with disease.” Controlled interpretation: “screening result positive; diagnostic evaluation may be needed.” This is not a style preference. Positive is a test interpretation, not necessarily a confirmed disease state. Public text may induce unwarranted certainty or anxiety. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

22A. Public-facing translation and human decision

Consider a web form presented alongside a next-step button. The source label is “screen positive”; an apparently natural localised phrase would assert “diagnosed with disease”. The approved meaning to carry is “screening result positive; diagnostic evaluation may be needed”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, public text may induce unwarranted certainty or anxiety. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Is disease confirmation asserted?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

22B. Machine-readable code and data-mapping test

Trace the record through a batch migration in which some values are absent. The relevant mapping inventory is screening_interpretation distinct from diagnosis. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “diagnosed with disease” masquerading as “screen positive”; the accepted reading remains “screening result positive; diagnostic evaluation may be needed”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

22C. Version, correction and post-release verification

Imagine a fallback string appearing before the new locale is fully approved. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Positive is a test interpretation, not necessarily a confirmed disease state. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. Public text may induce unwarranted certainty or anxiety. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Is disease confirmation asserted?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for screen-positive signal: Work independently in two roles. The first reviewer is given only “screen positive” and must write its weakest warranted assertion; the second reviewer is given only “screening result positive; diagnostic evaluation may be needed” in the target language and must do the same. Compare the two assertions, then introduce the tempting “diagnosed with disease” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

23. False-positive rate: three-view fidelity diagnostic

Source record: “false-positive rate among those without the condition.” Misleading candidate: “chance that a positive result is false.” Controlled interpretation: “false-positive rate using its specified denominator.” This is not a style preference. False-positive rate and the proportion of positive results that are false have different denominators. Counselling figures may mislead even when numerically unchanged. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

23A. Public-facing translation and human decision

Imagine a reminder sent through a channel that does not display the full underlying code. The source label is “false-positive rate among those without the condition”; an apparently natural localised phrase would assert “chance that a positive result is false”. The approved meaning to carry is “false-positive rate using its specified denominator”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, counselling figures may mislead even when numerically unchanged. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Can the denominator be reconstructed?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

23B. Machine-readable code and data-mapping test

Follow the field through an audit log and a derived analytics view. The relevant mapping inventory is metric_numerator, metric_denominator and population. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “chance that a positive result is false” masquerading as “false-positive rate among those without the condition”; the accepted reading remains “false-positive rate using its specified denominator”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

23C. Version, correction and post-release verification

Suppose an integration test passes but a downstream recipient reports a mismatch. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. False-positive rate and the proportion of positive results that are false have different denominators. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. Counselling figures may mislead even when numerically unchanged. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Can the denominator be reconstructed?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for false-positive rate: Work independently in two roles. The first reviewer is given only “false-positive rate among those without the condition” and must write its weakest warranted assertion; the second reviewer is given only “false-positive rate using its specified denominator” in the target language and must do the same. Compare the two assertions, then introduce the tempting “chance that a positive result is false” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

24. Sensitivity statement: three-view fidelity diagnostic

Source record: “sensitivity in the evaluated sample.” Misleading candidate: “probability the person has the condition.” Controlled interpretation: “sensitivity for the reported study and definition.” This is not a style preference. Sensitivity is not a personal post-test probability. A technical metric becomes misleading personalised risk. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

24A. Public-facing translation and human decision

Imagine a patient-facing portal after an automated screening workflow has produced a status message. The source label is “sensitivity in the evaluated sample”; an apparently natural localised phrase would assert “probability the person has the condition”. The approved meaning to carry is “sensitivity for the reported study and definition”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, a technical metric becomes misleading personalised risk. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Is the population and conditional direction explicit?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

24B. Machine-readable code and data-mapping test

Now follow the same record into an interoperability message handled by software rather than a reader. The relevant mapping inventory is metric_type, study_context and confidence_interval. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “probability the person has the condition” masquerading as “sensitivity in the evaluated sample”; the accepted reading remains “sensitivity for the reported study and definition”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

24C. Version, correction and post-release verification

Finally imagine a new terminology or software release affecting this field. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Sensitivity is not a personal post-test probability. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. A technical metric becomes misleading personalised risk. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Is the population and conditional direction explicit?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for sensitivity statement: Work independently in two roles. The first reviewer is given only “sensitivity in the evaluated sample” and must write its weakest warranted assertion; the second reviewer is given only “sensitivity for the reported study and definition” in the target language and must do the same. Compare the two assertions, then introduce the tempting “probability the person has the condition” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

25. Specificity statement: three-view fidelity diagnostic

Source record: “specificity in a validation cohort.” Misleading candidate: “certainty of a negative result.” Controlled interpretation: “specificity within the stated validation conditions.” This is not a style preference. Specificity concerns classification among those without a condition; it is not a general guarantee. The reader may disregard residual uncertainty. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

25A. Public-facing translation and human decision

Consider a plain-language explanation on an invitation page, where the recipient has limited context. The source label is “specificity in a validation cohort”; an apparently natural localised phrase would assert “certainty of a negative result”. The approved meaning to carry is “specificity within the stated validation conditions”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, the reader may disregard residual uncertainty. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Is the measure described by the correct condition?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

25B. Machine-readable code and data-mapping test

Move the example from a public screen to an export destined for a reporting system. The relevant mapping inventory is metric_type and cohort_definition. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “certainty of a negative result” masquerading as “specificity in a validation cohort”; the accepted reading remains “specificity within the stated validation conditions”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

25C. Version, correction and post-release verification

Now suppose a translator proposes a better term after the interface has gone live. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Specificity concerns classification among those without a condition; it is not a general guarantee. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. The reader may disregard residual uncertainty. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Is the measure described by the correct condition?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for specificity statement: Work independently in two roles. The first reviewer is given only “specificity in a validation cohort” and must write its weakest warranted assertion; the second reviewer is given only “specificity within the stated validation conditions” in the target language and must do the same. Compare the two assertions, then introduce the tempting “certainty of a negative result” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

26. Positive predictive value: three-view fidelity diagnostic

Source record: “positive predictive value at stated prevalence.” Misleading candidate: “constant property across all populations.” Controlled interpretation: “positive predictive value under the evaluated prevalence and setting.” This is not a style preference. Predictive value depends on prevalence and study design. A transferred percentage may appear portable when it is not. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

26A. Public-facing translation and human decision

Picture a multilingual dashboard whose short label must fit in a compact interface. The source label is “positive predictive value at stated prevalence”; an apparently natural localised phrase would assert “constant property across all populations”. The approved meaning to carry is “positive predictive value under the evaluated prevalence and setting”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, a transferred percentage may appear portable when it is not. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Does the receiving population match the estimate?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

26B. Machine-readable code and data-mapping test

Next test the field in a machine-to-machine exchange during reconciliation. The relevant mapping inventory is prevalence_context and evaluated_population. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “constant property across all populations” masquerading as “positive predictive value at stated prevalence”; the accepted reading remains “positive predictive value under the evaluated prevalence and setting”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

26C. Version, correction and post-release verification

Consider a complaint claiming that the displayed wording led to a different next action. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Predictive value depends on prevalence and study design. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. A transferred percentage may appear portable when it is not. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Does the receiving population match the estimate?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for positive predictive value: Work independently in two roles. The first reviewer is given only “positive predictive value at stated prevalence” and must write its weakest warranted assertion; the second reviewer is given only “positive predictive value under the evaluated prevalence and setting” in the target language and must do the same. Compare the two assertions, then introduce the tempting “constant property across all populations” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

27. Risk category: three-view fidelity diagnostic

Source record: “elevated risk group.” Misleading candidate: “disease present.” Controlled interpretation: “risk classification for screening eligibility.” This is not a style preference. Risk stratification is not a diagnosis and may be based on a specified rule. People may confuse future likelihood with current disease. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

27A. Public-facing translation and human decision

Consider a notification that a non-specialist reads before deciding whether to contact a service. The source label is “elevated risk group”; an apparently natural localised phrase would assert “disease present”. The approved meaning to carry is “risk classification for screening eligibility”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, people may confuse future likelihood with current disease. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “What outcome and time horizon does risk refer to?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

27B. Machine-readable code and data-mapping test

Consider the same state inside a data warehouse that computes aggregate screening counts. The relevant mapping inventory is risk_model_version and category_definition. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “disease present” masquerading as “elevated risk group”; the accepted reading remains “risk classification for screening eligibility”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

27C. Version, correction and post-release verification

Imagine a controlled rollback while older records remain available. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Risk stratification is not a diagnosis and may be based on a specified rule. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. People may confuse future likelihood with current disease. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “What outcome and time horizon does risk refer to?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for risk category: Work independently in two roles. The first reviewer is given only “elevated risk group” and must write its weakest warranted assertion; the second reviewer is given only “risk classification for screening eligibility” in the target language and must do the same. Compare the two assertions, then introduce the tempting “disease present” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

28. Family-history answer: three-view fidelity diagnostic

Source record: “family history unknown.” Misleading candidate: “no family history.” Controlled interpretation: “family history not known or not answered.” This is not a style preference. Unknown, not asked and no are three separate epistemic states. Eligibility rules may use an unjustified negative. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

28A. Public-facing translation and human decision

Imagine a telephone support agent reading a translated workflow message to a participant. The source label is “family history unknown”; an apparently natural localised phrase would assert “no family history”. The approved meaning to carry is “family history not known or not answered”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, eligibility rules may use an unjustified negative. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Was the question asked and answered?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

28B. Machine-readable code and data-mapping test

Now imagine a terminology service feeding a multilingual selection list. The relevant mapping inventory is answer_status and response_value. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “no family history” masquerading as “family history unknown”; the accepted reading remains “family history not known or not answered”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

28C. Version, correction and post-release verification

Suppose two language variants receive the new asset at different moments. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Unknown, not asked and no are three separate epistemic states. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. Eligibility rules may use an unjustified negative. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Was the question asked and answered?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for family-history answer: Work independently in two roles. The first reviewer is given only “family history unknown” and must write its weakest warranted assertion; the second reviewer is given only “family history not known or not answered” in the target language and must do the same. Compare the two assertions, then introduce the tempting “no family history” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

29. Symptom report: three-view fidelity diagnostic

Source record: “no symptoms reported.” Misleading candidate: “no symptoms exist.” Controlled interpretation: “no symptoms documented at this encounter.” This is not a style preference. A record of reporting is not proof of an absolute absence. A summary may conceal the limits of the observation. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

29A. Public-facing translation and human decision

Picture a printed summary that must remain intelligible long after the original encounter. The source label is “no symptoms reported”; an apparently natural localised phrase would assert “no symptoms exist”. The approved meaning to carry is “no symptoms documented at this encounter”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, a summary may conceal the limits of the observation. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Who observed or reported the finding?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

29B. Machine-readable code and data-mapping test

Next review the value inside a structured report copied across two organisations. The relevant mapping inventory is reporting_source and encounter_time. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “no symptoms exist” masquerading as “no symptoms reported”; the accepted reading remains “no symptoms documented at this encounter”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

29C. Version, correction and post-release verification

Consider a reviewer finding that the source data dictionary was underspecified. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. A record of reporting is not proof of an absolute absence. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. A summary may conceal the limits of the observation. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Who observed or reported the finding?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for symptom report: Work independently in two roles. The first reviewer is given only “no symptoms reported” and must write its weakest warranted assertion; the second reviewer is given only “no symptoms documented at this encounter” in the target language and must do the same. Compare the two assertions, then introduce the tempting “no symptoms exist” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

30. Recall due date: three-view fidelity diagnostic

Source record: “recall due next month.” Misleading candidate: “appointment confirmed next month.” Controlled interpretation: “recall becomes due next month; booking still required.” This is not a style preference. A due date sets an action target, not a reserved appointment. A participant may miss the actual booking process. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

30A. Public-facing translation and human decision

Consider a web form presented alongside a next-step button. The source label is “recall due next month”; an apparently natural localised phrase would assert “appointment confirmed next month”. The approved meaning to carry is “recall becomes due next month; booking still required”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, a participant may miss the actual booking process. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Is there a confirmed appointment identifier?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

30B. Machine-readable code and data-mapping test

Trace the record through a batch migration in which some values are absent. The relevant mapping inventory is recall_due and appointment_id. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “appointment confirmed next month” masquerading as “recall due next month”; the accepted reading remains “recall becomes due next month; booking still required”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

30C. Version, correction and post-release verification

Imagine a fallback string appearing before the new locale is fully approved. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. A due date sets an action target, not a reserved appointment. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. A participant may miss the actual booking process. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Is there a confirmed appointment identifier?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for recall due date: Work independently in two roles. The first reviewer is given only “recall due next month” and must write its weakest warranted assertion; the second reviewer is given only “recall becomes due next month; booking still required” in the target language and must do the same. Compare the two assertions, then introduce the tempting “appointment confirmed next month” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

31. Referral issued: three-view fidelity diagnostic

Source record: “referral sent.” Misleading candidate: “specialist accepted referral.” Controlled interpretation: “referral submitted; acceptance pending.” This is not a style preference. Sending information does not prove the receiving service accepted responsibility. A pathway may be falsely marked closed. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

31A. Public-facing translation and human decision

Imagine a reminder sent through a channel that does not display the full underlying code. The source label is “referral sent”; an apparently natural localised phrase would assert “specialist accepted referral”. The approved meaning to carry is “referral submitted; acceptance pending”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, a pathway may be falsely marked closed. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Who accepted the referral?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

31B. Machine-readable code and data-mapping test

Follow the field through an audit log and a derived analytics view. The relevant mapping inventory is referral_sent_at, accepted_at and receiving_party. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “specialist accepted referral” masquerading as “referral sent”; the accepted reading remains “referral submitted; acceptance pending”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

31C. Version, correction and post-release verification

Suppose an integration test passes but a downstream recipient reports a mismatch. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Sending information does not prove the receiving service accepted responsibility. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. A pathway may be falsely marked closed. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Who accepted the referral?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for referral issued: Work independently in two roles. The first reviewer is given only “referral sent” and must write its weakest warranted assertion; the second reviewer is given only “referral submitted; acceptance pending” in the target language and must do the same. Compare the two assertions, then introduce the tempting “specialist accepted referral” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

32. Follow-up closure: three-view fidelity diagnostic

Source record: “no outstanding follow-up task in this system.” Misleading candidate: “follow-up clinically complete.” Controlled interpretation: “no open local task; cross-system completion not established.” This is not a style preference. Local workflow closure is narrower than clinical completion across organisations. Audit statistics may overstate completed care. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

32A. Public-facing translation and human decision

Imagine a patient-facing portal after an automated screening workflow has produced a status message. The source label is “no outstanding follow-up task in this system”; an apparently natural localised phrase would assert “follow-up clinically complete”. The approved meaning to carry is “no open local task; cross-system completion not established”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, audit statistics may overstate completed care. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Which record proves the whole pathway ended?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

32B. Machine-readable code and data-mapping test

Now follow the same record into an interoperability message handled by software rather than a reader. The relevant mapping inventory is closure_reason and evidence_scope. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “follow-up clinically complete” masquerading as “no outstanding follow-up task in this system”; the accepted reading remains “no open local task; cross-system completion not established”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

32C. Version, correction and post-release verification

Finally imagine a new terminology or software release affecting this field. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Local workflow closure is narrower than clinical completion across organisations. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. Audit statistics may overstate completed care. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Which record proves the whole pathway ended?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for follow-up closure: Work independently in two roles. The first reviewer is given only “no outstanding follow-up task in this system” and must write its weakest warranted assertion; the second reviewer is given only “no open local task; cross-system completion not established” in the target language and must do the same. Compare the two assertions, then introduce the tempting “follow-up clinically complete” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

33. Human review: three-view fidelity diagnostic

Source record: “review requested.” Misleading candidate: “human review completed.” Controlled interpretation: “review requested and awaiting decision.” This is not a style preference. Escalation and adjudication are different statuses. An algorithmic result may be treated as independently confirmed. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

33A. Public-facing translation and human decision

Consider a plain-language explanation on an invitation page, where the recipient has limited context. The source label is “review requested”; an apparently natural localised phrase would assert “human review completed”. The approved meaning to carry is “review requested and awaiting decision”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, an algorithmic result may be treated as independently confirmed. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Where is the actual review decision?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

33B. Machine-readable code and data-mapping test

Move the example from a public screen to an export destined for a reporting system. The relevant mapping inventory is review_request_id and reviewer_signoff. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “human review completed” masquerading as “review requested”; the accepted reading remains “review requested and awaiting decision”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

33C. Version, correction and post-release verification

Now suppose a translator proposes a better term after the interface has gone live. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Escalation and adjudication are different statuses. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. An algorithmic result may be treated as independently confirmed. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Where is the actual review decision?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for human review: Work independently in two roles. The first reviewer is given only “review requested” and must write its weakest warranted assertion; the second reviewer is given only “review requested and awaiting decision” in the target language and must do the same. Compare the two assertions, then introduce the tempting “human review completed” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

34. Decision override: three-view fidelity diagnostic

Source record: “authorised override logged.” Misleading candidate: “source result deleted.” Controlled interpretation: “override recorded separately from the original output.” This is not a style preference. Overrides should preserve an accountable history rather than erase a prior signal. Investigators could lose evidence about the decision process. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

34A. Public-facing translation and human decision

Picture a multilingual dashboard whose short label must fit in a compact interface. The source label is “authorised override logged”; an apparently natural localised phrase would assert “source result deleted”. The approved meaning to carry is “override recorded separately from the original output”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, investigators could lose evidence about the decision process. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Can the original output still be reconstructed?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

34B. Machine-readable code and data-mapping test

Next test the field in a machine-to-machine exchange during reconciliation. The relevant mapping inventory is original_result, override_reason and actor. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “source result deleted” masquerading as “authorised override logged”; the accepted reading remains “override recorded separately from the original output”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

34C. Version, correction and post-release verification

Consider a complaint claiming that the displayed wording led to a different next action. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Overrides should preserve an accountable history rather than erase a prior signal. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. Investigators could lose evidence about the decision process. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Can the original output still be reconstructed?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for decision override: Work independently in two roles. The first reviewer is given only “authorised override logged” and must write its weakest warranted assertion; the second reviewer is given only “override recorded separately from the original output” in the target language and must do the same. Compare the two assertions, then introduce the tempting “source result deleted” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

35. Correction state: three-view fidelity diagnostic

Source record: “corrected label queued.” Misleading candidate: “corrected label published everywhere.” Controlled interpretation: “correction approved; distribution verification pending.” This is not a style preference. Approval and end-user deployment are separate release stages. Some languages could continue to display superseded text. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

35A. Public-facing translation and human decision

Consider a notification that a non-specialist reads before deciding whether to contact a service. The source label is “corrected label queued”; an apparently natural localised phrase would assert “corrected label published everywhere”. The approved meaning to carry is “correction approved; distribution verification pending”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, some languages could continue to display superseded text. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Which release actually reached users?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

35B. Machine-readable code and data-mapping test

Consider the same state inside a data warehouse that computes aggregate screening counts. The relevant mapping inventory is approval_state, deployed_locales and verification_receipts. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “corrected label published everywhere” masquerading as “corrected label queued”; the accepted reading remains “correction approved; distribution verification pending”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

35C. Version, correction and post-release verification

Imagine a controlled rollback while older records remain available. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Approval and end-user deployment are separate release stages. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. Some languages could continue to display superseded text. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Which release actually reached users?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for correction state: Work independently in two roles. The first reviewer is given only “corrected label queued” and must write its weakest warranted assertion; the second reviewer is given only “correction approved; distribution verification pending” in the target language and must do the same. Compare the two assertions, then introduce the tempting “corrected label published everywhere” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

36. Concept identifier: three-view fidelity diagnostic

Source record: “stable concept identifier.” Misleading candidate: “translated phrase as the primary machine key.” Controlled interpretation: “retain stable concept identifier; translate only approved human display.” This is not a style preference. A human-language label is not a safe substitute for a terminology identifier. Records may become incomparable after changing labels. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

36A. Public-facing translation and human decision

Imagine a telephone support agent reading a translated workflow message to a participant. The source label is “stable concept identifier”; an apparently natural localised phrase would assert “translated phrase as the primary machine key”. The approved meaning to carry is “retain stable concept identifier; translate only approved human display”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, records may become incomparable after changing labels. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Does a language switch leave the concept code unchanged?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

36B. Machine-readable code and data-mapping test

Now imagine a terminology service feeding a multilingual selection list. The relevant mapping inventory is Coding.system, Coding.code and display. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “translated phrase as the primary machine key” masquerading as “stable concept identifier”; the accepted reading remains “retain stable concept identifier; translate only approved human display”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

36C. Version, correction and post-release verification

Suppose two language variants receive the new asset at different moments. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. A human-language label is not a safe substitute for a terminology identifier. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. Records may become incomparable after changing labels. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Does a language switch leave the concept code unchanged?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for concept identifier: Work independently in two roles. The first reviewer is given only “stable concept identifier” and must write its weakest warranted assertion; the second reviewer is given only “retain stable concept identifier; translate only approved human display” in the target language and must do the same. Compare the two assertions, then introduce the tempting “translated phrase as the primary machine key” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

37. Coding system: three-view fidelity diagnostic

Source record: “same visible code string.” Misleading candidate: “same concept across any code system.” Controlled interpretation: “code interpreted only with its named code system.” This is not a style preference. An identical alphanumeric code can mean different things in different namespaces. Data may silently merge different categories. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

37A. Public-facing translation and human decision

Picture a printed summary that must remain intelligible long after the original encounter. The source label is “same visible code string”; an apparently natural localised phrase would assert “same concept across any code system”. The approved meaning to carry is “code interpreted only with its named code system”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, data may silently merge different categories. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Is the code system URI present?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

37B. Machine-readable code and data-mapping test

Next review the value inside a structured report copied across two organisations. The relevant mapping inventory is system URI plus code plus version if applicable. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “same concept across any code system” masquerading as “same visible code string”; the accepted reading remains “code interpreted only with its named code system”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

37C. Version, correction and post-release verification

Consider a reviewer finding that the source data dictionary was underspecified. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. An identical alphanumeric code can mean different things in different namespaces. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. Data may silently merge different categories. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Is the code system URI present?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for coding system: Work independently in two roles. The first reviewer is given only “same visible code string” and must write its weakest warranted assertion; the second reviewer is given only “code interpreted only with its named code system” in the target language and must do the same. Compare the two assertions, then introduce the tempting “same concept across any code system” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

38. Preferred synonym: three-view fidelity diagnostic

Source record: “preferred term for one dialect.” Misleading candidate: “universally preferred term.” Controlled interpretation: “preferred display under the specified language reference set.” This is not a style preference. A term preferred in one locale can be merely acceptable in another. A valid synonym may be incorrectly rejected or enforced. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

38A. Public-facing translation and human decision

Consider a web form presented alongside a next-step button. The source label is “preferred term for one dialect”; an apparently natural localised phrase would assert “universally preferred term”. The approved meaning to carry is “preferred display under the specified language reference set”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, a valid synonym may be incorrectly rejected or enforced. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Which language reference set establishes preference?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

38B. Machine-readable code and data-mapping test

Trace the record through a batch migration in which some values are absent. The relevant mapping inventory is conceptId, descriptionId and language refset. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “universally preferred term” masquerading as “preferred term for one dialect”; the accepted reading remains “preferred display under the specified language reference set”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

38C. Version, correction and post-release verification

Imagine a fallback string appearing before the new locale is fully approved. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. A term preferred in one locale can be merely acceptable in another. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. A valid synonym may be incorrectly rejected or enforced. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Which language reference set establishes preference?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for preferred synonym: Work independently in two roles. The first reviewer is given only “preferred term for one dialect” and must write its weakest warranted assertion; the second reviewer is given only “preferred display under the specified language reference set” in the target language and must do the same. Compare the two assertions, then introduce the tempting “universally preferred term” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

39. Display and designation: three-view fidelity diagnostic

Source record: “consumer-facing display label.” Misleading candidate: “the only authorised term for every setting.” Controlled interpretation: “contextual display; retain approved designations and source concept.” This is not a style preference. A short UI label may differ from a formal description without changing code identity. Documentation and screens may diverge without governance. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

39A. Public-facing translation and human decision

Imagine a reminder sent through a channel that does not display the full underlying code. The source label is “consumer-facing display label”; an apparently natural localised phrase would assert “the only authorised term for every setting”. The approved meaning to carry is “contextual display; retain approved designations and source concept”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, documentation and screens may diverge without governance. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Is this an authorised display for this audience?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

39B. Machine-readable code and data-mapping test

Follow the field through an audit log and a derived analytics view. The relevant mapping inventory is display, designation, language and intended use. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “the only authorised term for every setting” masquerading as “consumer-facing display label”; the accepted reading remains “contextual display; retain approved designations and source concept”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

39C. Version, correction and post-release verification

Suppose an integration test passes but a downstream recipient reports a mismatch. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. A short UI label may differ from a formal description without changing code identity. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. Documentation and screens may diverge without governance. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Is this an authorised display for this audience?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for display and designation: Work independently in two roles. The first reviewer is given only “consumer-facing display label” and must write its weakest warranted assertion; the second reviewer is given only “contextual display; retain approved designations and source concept” in the target language and must do the same. Compare the two assertions, then introduce the tempting “the only authorised term for every setting” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

40. Free-text nuance: three-view fidelity diagnostic

Source record: “patient said ‘not sure’.” Misleading candidate: “coded as no.” Controlled interpretation: “patient’s uncertainty retained as text pending coding review.” This is not a style preference. Free text can preserve nuance that a selected category does not express. A patient quotation may become an incorrect categorical answer. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

40A. Public-facing translation and human decision

Imagine a patient-facing portal after an automated screening workflow has produced a status message. The source label is “patient said ‘not sure’”; an apparently natural localised phrase would assert “coded as no”. The approved meaning to carry is “patient’s uncertainty retained as text pending coding review”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, a patient quotation may become an incorrect categorical answer. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Was the original voice retained?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

40B. Machine-readable code and data-mapping test

Now follow the same record into an interoperability message handled by software rather than a reader. The relevant mapping inventory is CodeableConcept.text and separately reviewed coding. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “coded as no” masquerading as “patient said ‘not sure’”; the accepted reading remains “patient’s uncertainty retained as text pending coding review”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

40C. Version, correction and post-release verification

Finally imagine a new terminology or software release affecting this field. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Free text can preserve nuance that a selected category does not express. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. A patient quotation may become an incorrect categorical answer. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Was the original voice retained?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for free-text nuance: Work independently in two roles. The first reviewer is given only “patient said ‘not sure’” and must write its weakest warranted assertion; the second reviewer is given only “patient’s uncertainty retained as text pending coding review” in the target language and must do the same. Compare the two assertions, then introduce the tempting “coded as no” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

41. Value-set membership: three-view fidelity diagnostic

Source record: “value set expanded yesterday.” Misleading candidate: “membership fixed for all time.” Controlled interpretation: “membership checked against a named value-set version and expansion.” This is not a style preference. A value set is a governed collection; an expansion can change with versions and parameters. Dropdown choices may drift without review. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

41A. Public-facing translation and human decision

Consider a plain-language explanation on an invitation page, where the recipient has limited context. The source label is “value set expanded yesterday”; an apparently natural localised phrase would assert “membership fixed for all time”. The approved meaning to carry is “membership checked against a named value-set version and expansion”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, dropdown choices may drift without review. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Can the expansion be reproduced?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

41B. Machine-readable code and data-mapping test

Move the example from a public screen to an export destined for a reporting system. The relevant mapping inventory is canonical URL, version, expansion parameters and timestamp. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “membership fixed for all time” masquerading as “value set expanded yesterday”; the accepted reading remains “membership checked against a named value-set version and expansion”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

41C. Version, correction and post-release verification

Now suppose a translator proposes a better term after the interface has gone live. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. A value set is a governed collection; an expansion can change with versions and parameters. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. Dropdown choices may drift without review. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Can the expansion be reproduced?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for value-set membership: Work independently in two roles. The first reviewer is given only “value set expanded yesterday” and must write its weakest warranted assertion; the second reviewer is given only “membership checked against a named value-set version and expansion” in the target language and must do the same. Compare the two assertions, then introduce the tempting “membership fixed for all time” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

42. Value-set exclusion: three-view fidelity diagnostic

Source record: “code excluded by this rule.” Misleading candidate: “code invalid globally.” Controlled interpretation: “code excluded from this specific permitted selection set.” This is not a style preference. Exclusion from one workflow’s value set does not invalidate a concept everywhere. A user may wrongly treat valid information as impossible. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

42A. Public-facing translation and human decision

Picture a multilingual dashboard whose short label must fit in a compact interface. The source label is “code excluded by this rule”; an apparently natural localised phrase would assert “code invalid globally”. The approved meaning to carry is “code excluded from this specific permitted selection set”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, a user may wrongly treat valid information as impossible. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Does exclusion apply only to this workflow?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

42B. Machine-readable code and data-mapping test

Next test the field in a machine-to-machine exchange during reconciliation. The relevant mapping inventory is value_set_uri and exclusion_rule. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “code invalid globally” masquerading as “code excluded by this rule”; the accepted reading remains “code excluded from this specific permitted selection set”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

42C. Version, correction and post-release verification

Consider a complaint claiming that the displayed wording led to a different next action. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Exclusion from one workflow’s value set does not invalidate a concept everywhere. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. A user may wrongly treat valid information as impossible. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Does exclusion apply only to this workflow?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for value-set exclusion: Work independently in two roles. The first reviewer is given only “code excluded by this rule” and must write its weakest warranted assertion; the second reviewer is given only “code excluded from this specific permitted selection set” in the target language and must do the same. Compare the two assertions, then introduce the tempting “code invalid globally” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

43. Terminology retirement: three-view fidelity diagnostic

Source record: “term inactive after release.” Misleading candidate: “clinical concept never existed.” Controlled interpretation: “term inactive in the named release; historical records preserved.” This is not a style preference. Retirement controls future use; historical evidence must not be rewritten as if absent. Legacy records may be silently distorted. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

43A. Public-facing translation and human decision

Consider a notification that a non-specialist reads before deciding whether to contact a service. The source label is “term inactive after release”; an apparently natural localised phrase would assert “clinical concept never existed”. The approved meaning to carry is “term inactive in the named release; historical records preserved”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, legacy records may be silently distorted. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “How are earlier records rendered?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

43B. Machine-readable code and data-mapping test

Consider the same state inside a data warehouse that computes aggregate screening counts. The relevant mapping inventory is effective_date, inactive flag and replacement mapping. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “clinical concept never existed” masquerading as “term inactive after release”; the accepted reading remains “term inactive in the named release; historical records preserved”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

43C. Version, correction and post-release verification

Imagine a controlled rollback while older records remain available. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Retirement controls future use; historical evidence must not be rewritten as if absent. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. Legacy records may be silently distorted. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “How are earlier records rendered?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for terminology retirement: Work independently in two roles. The first reviewer is given only “term inactive after release” and must write its weakest warranted assertion; the second reviewer is given only “term inactive in the named release; historical records preserved” in the target language and must do the same. Compare the two assertions, then introduce the tempting “clinical concept never existed” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

44. Code mapping relation: three-view fidelity diagnostic

Source record: “source concept mapped approximately.” Misleading candidate: “source and target concepts identical.” Controlled interpretation: “mapping qualified as approximate and subject to review.” This is not a style preference. Cross-system mappings can be narrower, broader or inexact. A receiving system may overclaim equivalence. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

44A. Public-facing translation and human decision

Imagine a telephone support agent reading a translated workflow message to a participant. The source label is “source concept mapped approximately”; an apparently natural localised phrase would assert “source and target concepts identical”. The approved meaning to carry is “mapping qualified as approximate and subject to review”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, a receiving system may overclaim equivalence. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Is mapping equivalence explicitly justified?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

44B. Machine-readable code and data-mapping test

Now imagine a terminology service feeding a multilingual selection list. The relevant mapping inventory is ConceptMap relationship and rationale. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “source and target concepts identical” masquerading as “source concept mapped approximately”; the accepted reading remains “mapping qualified as approximate and subject to review”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

44C. Version, correction and post-release verification

Suppose two language variants receive the new asset at different moments. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Cross-system mappings can be narrower, broader or inexact. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. A receiving system may overclaim equivalence. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Is mapping equivalence explicitly justified?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for code mapping relation: Work independently in two roles. The first reviewer is given only “source concept mapped approximately” and must write its weakest warranted assertion; the second reviewer is given only “mapping qualified as approximate and subject to review” in the target language and must do the same. Compare the two assertions, then introduce the tempting “source and target concepts identical” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

45. Release version: three-view fidelity diagnostic

Source record: “release 3.2 effective next cycle.” Misleading candidate: “release 3.2 already active.” Controlled interpretation: “release 3.2 scheduled; prior release remains operative until activation.” This is not a style preference. Publication date, effective date and actual deployment are not one event. Staff may follow instructions from a premature release. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

45A. Public-facing translation and human decision

Picture a printed summary that must remain intelligible long after the original encounter. The source label is “release 3.2 effective next cycle”; an apparently natural localised phrase would assert “release 3.2 already active”. The approved meaning to carry is “release 3.2 scheduled; prior release remains operative until activation”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, staff may follow instructions from a premature release. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Which version governs this encounter?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

45B. Machine-readable code and data-mapping test

Next review the value inside a structured report copied across two organisations. The relevant mapping inventory is published_at, effective_at and deployed_at. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “release 3.2 already active” masquerading as “release 3.2 effective next cycle”; the accepted reading remains “release 3.2 scheduled; prior release remains operative until activation”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

45C. Version, correction and post-release verification

Consider a reviewer finding that the source data dictionary was underspecified. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Publication date, effective date and actual deployment are not one event. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. Staff may follow instructions from a premature release. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Which version governs this encounter?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for release version: Work independently in two roles. The first reviewer is given only “release 3.2 effective next cycle” and must write its weakest warranted assertion; the second reviewer is given only “release 3.2 scheduled; prior release remains operative until activation” in the target language and must do the same. Compare the two assertions, then introduce the tempting “release 3.2 already active” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

46. Fallback language: three-view fidelity diagnostic

Source record: “unreviewed fallback shown.” Misleading candidate: “approved translation.” Controlled interpretation: “fallback from source language; translation not yet validated.” This is not a style preference. A fallback can serve continuity but must not impersonate a reviewed localisation. Readers could trust an unreviewed safety-critical sentence. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

46A. Public-facing translation and human decision

Consider a web form presented alongside a next-step button. The source label is “unreviewed fallback shown”; an apparently natural localised phrase would assert “approved translation”. The approved meaning to carry is “fallback from source language; translation not yet validated”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, readers could trust an unreviewed safety-critical sentence. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Which language was actually delivered?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

46B. Machine-readable code and data-mapping test

Trace the record through a batch migration in which some values are absent. The relevant mapping inventory is requested_locale, served_locale and translation_review_state. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “approved translation” masquerading as “unreviewed fallback shown”; the accepted reading remains “fallback from source language; translation not yet validated”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

46C. Version, correction and post-release verification

Imagine a fallback string appearing before the new locale is fully approved. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. A fallback can serve continuity but must not impersonate a reviewed localisation. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. Readers could trust an unreviewed safety-critical sentence. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Which language was actually delivered?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for fallback language: Work independently in two roles. The first reviewer is given only “unreviewed fallback shown” and must write its weakest warranted assertion; the second reviewer is given only “fallback from source language; translation not yet validated” in the target language and must do the same. Compare the two assertions, then introduce the tempting “approved translation” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

47. Machine translation confidence: three-view fidelity diagnostic

Source record: “machine candidate pending human review.” Misleading candidate: “professionally verified translation.” Controlled interpretation: “machine-generated candidate not yet approved.” This is not a style preference. Model confidence or fluent output is not evidence of clinical translation acceptance. A provisional message may be published as signed-off advice. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

47A. Public-facing translation and human decision

Imagine a reminder sent through a channel that does not display the full underlying code. The source label is “machine candidate pending human review”; an apparently natural localised phrase would assert “professionally verified translation”. The approved meaning to carry is “machine-generated candidate not yet approved”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, a provisional message may be published as signed-off advice. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Who performed the risk-appropriate review?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

47B. Machine-readable code and data-mapping test

Follow the field through an audit log and a derived analytics view. The relevant mapping inventory is translation_origin and approval_status. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “professionally verified translation” masquerading as “machine candidate pending human review”; the accepted reading remains “machine-generated candidate not yet approved”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

47C. Version, correction and post-release verification

Suppose an integration test passes but a downstream recipient reports a mismatch. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Model confidence or fluent output is not evidence of clinical translation acceptance. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. A provisional message may be published as signed-off advice. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Who performed the risk-appropriate review?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for machine translation confidence: Work independently in two roles. The first reviewer is given only “machine candidate pending human review” and must write its weakest warranted assertion; the second reviewer is given only “machine-generated candidate not yet approved” in the target language and must do the same. Compare the two assertions, then introduce the tempting “professionally verified translation” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

48. Audit log entry: three-view fidelity diagnostic

Source record: “reviewer changed text label.” Misleading candidate: “clinical result changed.” Controlled interpretation: “human-facing label revised; clinical observation unchanged.” This is not a style preference. Editing a display string must not create a false medical event. An audit could wrongly report altered results. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

48A. Public-facing translation and human decision

Imagine a patient-facing portal after an automated screening workflow has produced a status message. The source label is “reviewer changed text label”; an apparently natural localised phrase would assert “clinical result changed”. The approved meaning to carry is “human-facing label revised; clinical observation unchanged”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, an audit could wrongly report altered results. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Did any coded clinical value change?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

48B. Machine-readable code and data-mapping test

Now follow the same record into an interoperability message handled by software rather than a reader. The relevant mapping inventory is field_path, previous_label, new_label and actor. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “clinical result changed” masquerading as “reviewer changed text label”; the accepted reading remains “human-facing label revised; clinical observation unchanged”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

48C. Version, correction and post-release verification

Finally imagine a new terminology or software release affecting this field. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Editing a display string must not create a false medical event. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. An audit could wrongly report altered results. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Did any coded clinical value change?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for audit log entry: Work independently in two roles. The first reviewer is given only “reviewer changed text label” and must write its weakest warranted assertion; the second reviewer is given only “human-facing label revised; clinical observation unchanged” in the target language and must do the same. Compare the two assertions, then introduce the tempting “clinical result changed” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

49. De-identified export: three-view fidelity diagnostic

Source record: “identifiers removed from this extract.” Misleading candidate: “irreversibly anonymous.” Controlled interpretation: “direct identifiers removed; residual re-identification risk assessed separately.” This is not a style preference. Removing visible names is not proof that re-identification is impossible. A data-sharing notice may overpromise confidentiality. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

49A. Public-facing translation and human decision

Consider a plain-language explanation on an invitation page, where the recipient has limited context. The source label is “identifiers removed from this extract”; an apparently natural localised phrase would assert “irreversibly anonymous”. The approved meaning to carry is “direct identifiers removed; residual re-identification risk assessed separately”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, a data-sharing notice may overpromise confidentiality. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “What quasi-identifiers remain?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

49B. Machine-readable code and data-mapping test

Move the example from a public screen to an export destined for a reporting system. The relevant mapping inventory is deidentification_method and residual_risk_record. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “irreversibly anonymous” masquerading as “identifiers removed from this extract”; the accepted reading remains “direct identifiers removed; residual re-identification risk assessed separately”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

49C. Version, correction and post-release verification

Now suppose a translator proposes a better term after the interface has gone live. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Removing visible names is not proof that re-identification is impossible. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. A data-sharing notice may overpromise confidentiality. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “What quasi-identifiers remain?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for de-identified export: Work independently in two roles. The first reviewer is given only “identifiers removed from this extract” and must write its weakest warranted assertion; the second reviewer is given only “direct identifiers removed; residual re-identification risk assessed separately” in the target language and must do the same. Compare the two assertions, then introduce the tempting “irreversibly anonymous” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

50. Disclosure authority: three-view fidelity diagnostic

Source record: “dataset available to approved analysts.” Misleading candidate: “open public dataset.” Controlled interpretation: “data accessible only to authorised roles under the stated terms.” This is not a style preference. Permission to analyse does not automatically permit public distribution. Sensitive information may be disclosed too broadly. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

50A. Public-facing translation and human decision

Picture a multilingual dashboard whose short label must fit in a compact interface. The source label is “dataset available to approved analysts”; an apparently natural localised phrase would assert “open public dataset”. The approved meaning to carry is “data accessible only to authorised roles under the stated terms”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, sensitive information may be disclosed too broadly. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Who may access the resource for what purpose?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

50B. Machine-readable code and data-mapping test

Next test the field in a machine-to-machine exchange during reconciliation. The relevant mapping inventory is access_policy_id and role scope. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “open public dataset” masquerading as “dataset available to approved analysts”; the accepted reading remains “data accessible only to authorised roles under the stated terms”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

50C. Version, correction and post-release verification

Consider a complaint claiming that the displayed wording led to a different next action. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Permission to analyse does not automatically permit public distribution. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. Sensitive information may be disclosed too broadly. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Who may access the resource for what purpose?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for disclosure authority: Work independently in two roles. The first reviewer is given only “dataset available to approved analysts” and must write its weakest warranted assertion; the second reviewer is given only “data accessible only to authorised roles under the stated terms” in the target language and must do the same. Compare the two assertions, then introduce the tempting “open public dataset” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

51. Quality-control state: three-view fidelity diagnostic

Source record: “quality control passed.” Misleading candidate: “result clinically released.” Controlled interpretation: “analytical control accepted; result release is separate.” This is not a style preference. Quality checks on the process do not automatically authorize final reporting. A preliminary result may be treated as final. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

51A. Public-facing translation and human decision

Consider a notification that a non-specialist reads before deciding whether to contact a service. The source label is “quality control passed”; an apparently natural localised phrase would assert “result clinically released”. The approved meaning to carry is “analytical control accepted; result release is separate”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, a preliminary result may be treated as final. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “Is there a separate authorized release event?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

51B. Machine-readable code and data-mapping test

Consider the same state inside a data warehouse that computes aggregate screening counts. The relevant mapping inventory is qc_status and report_release_status. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “result clinically released” masquerading as “quality control passed”; the accepted reading remains “analytical control accepted; result release is separate”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

51C. Version, correction and post-release verification

Imagine a controlled rollback while older records remain available. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Quality checks on the process do not automatically authorize final reporting. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. A preliminary result may be treated as final. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “Is there a separate authorized release event?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for quality-control state: Work independently in two roles. The first reviewer is given only “quality control passed” and must write its weakest warranted assertion; the second reviewer is given only “analytical control accepted; result release is separate” in the target language and must do the same. Compare the two assertions, then introduce the tempting “result clinically released” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

52. Programme cohort: three-view fidelity diagnostic

Source record: “participants invited in this period.” Misleading candidate: “people screened in this period.” Controlled interpretation: “invited cohort measured by invitation event.” This is not a style preference. Invitation counts and completed screenings use different denominators. Performance claims can be inflated by replacing counts. The task for a translator is to preserve the truth conditions of the source: what event has occurred, whose evidence supports it and which stronger conclusion would be unwarranted.

52A. Public-facing translation and human decision

Imagine a telephone support agent reading a translated workflow message to a participant. The source label is “participants invited in this period”; an apparently natural localised phrase would assert “people screened in this period”. The approved meaning to carry is “invited cohort measured by invitation event”. A reviewer should first identify the decision the reader is likely to make, then compare whether that decision is licensed in both versions. In this case, performance claims can be inflated by replacing counts. The right wording must protect the difference without burying it in jargon or overconfident reassurance.

A useful revision sequence is to paraphrase the source in simpler language, mark every assertion that becomes stronger or weaker, translate the controlled paraphrase into the target language, and then ask an independent bilingual reviewer to reconstruct the original state from the target wording alone. The reviewer is not asked which version sounds nicer. The decisive prompt is: “What exactly was counted?” If the answer changes after translation, the meaning has changed even if the prose flows. Make the next action explicit only where the authorised source provides one; do not add an invented booking, diagnosis, permission or follow-up requirement.

52B. Machine-readable code and data-mapping test

Now imagine a terminology service feeding a multilingual selection list. The relevant mapping inventory is cohort_event and period boundaries. This must be specified with types, permissible values, provenance and the applicable source version before localisation. A data integrator may preserve the visible text while storing a wrong value, or store the right coded value while showing a misleading label. Those are independent faults. The acceptance test should examine both the machine identity and the human-facing output, including what happens when the source value is missing, outside a value set, or valid only for an earlier release.

Construct three fictional rows: a straightforward example, a boundary example and an unresolved-state example. For each row record the source field, code-system context where relevant, target display, expected action and evidence that authorises the classification. Round-trip through the receiving application and inspect whether it reconstructed the same state. The sentinel error for this case is “people screened in this period” masquerading as “participants invited in this period”; the accepted reading remains “invited cohort measured by invitation event”. If the data can no longer distinguish those two propositions, stop the release and repair the mapping at its source rather than hiding the error in a user-interface tooltip.

52C. Version, correction and post-release verification

Suppose two language variants receive the new asset at different moments. Before changing the asset, record its original dictionary revision, permitted value domain, translated display, review decision, effective date and application destination. Invitation counts and completed screenings use different denominators. An authorised correction should state exactly which assertion was wrong, which contexts are affected, which historical records must remain interpretable and whether a human-facing revision also needs a data-model change. Do not treat a terminology change as proof that a screening decision changed retrospectively.

The release checklist asks who approved the new wording, which language builds received it, whether the old version remains in caches or exports, and what a reader could mistakenly do while versions differ. Performance claims can be inflated by replacing counts. The deployment proof is not an email saying “updated”; it is a reproducible read of the actual served display, an accepted mapping test and a trace of the previous version. Finish by asking “What exactly was counted?” and keep the answer next to the evidence, not inside an undocumented editorial note.

Reviewer workshop for programme cohort: Work independently in two roles. The first reviewer is given only “participants invited in this period” and must write its weakest warranted assertion; the second reviewer is given only “invited cohort measured by invitation event” in the target language and must do the same. Compare the two assertions, then introduce the tempting “people screened in this period” wording as an adversarial test. Explain which field, actor, time or condition the adversarial wording changes. Record the fix as a testable invariant suitable for use again whenever this field appears in a consent form, message template, analytics export or future translation brief.

7. Cross-document workshops: checking one meaning through a whole workflow

Workshop 1: Invitation to attendance

Build a fictional record that passes sequentially through these states: Invitation generated → dispatch attempted → delivery evidenced → appointment booked → attendance recorded. Translate the label at each boundary into two target-language drafts, one intended for a professional and one for a lay reader. Preserve the same event identity while changing register. An invitation is not the booking; neither is proof of attendance. Ask which action could be taken after each stage and which action remains premature; these are the test oracles for acceptance.

Next create an adversarial version in which the second or third state is mistakenly labelled as the final one. Trace how that claim affects a dashboard total, a participant notification, a report, a help-desk explanation and an audit trail. Where the error changes the state of evidence, mark it as semantic rather than stylistic. Finally simulate a source revision: change only one authorised condition and identify every translated asset and structured field that must be retested. Write a receipt showing old and new dictionary versions, approval roles and deployment evidence.

Close the workshop with an independent reconstruction exercise. A reviewer who did not read the source must infer the state sequence, responsible actor, original uncertainty, version in use and permissible next action from the translation alone. Compare that reconstruction with the source record and record all differences. An apparently fluent candidate fails if it changes any of those answers. This test is useful because it measures actual decision equivalence rather than superficial word agreement.

Workshop 2: Eligibility to choice

Build a fictional record that passes sequentially through these states: Eligibility not evaluated → assessment started → eligible under current rule → offer communicated → consent decision recorded. Translate the label at each boundary into two target-language drafts, one intended for a professional and one for a lay reader. Preserve the same event identity while changing register. Eligibility does not itself confer consent. Ask which action could be taken after each stage and which action remains premature; these are the test oracles for acceptance.

Next create an adversarial version in which the second or third state is mistakenly labelled as the final one. Trace how that claim affects a dashboard total, a participant notification, a report, a help-desk explanation and an audit trail. Where the error changes the state of evidence, mark it as semantic rather than stylistic. Finally simulate a source revision: change only one authorised condition and identify every translated asset and structured field that must be retested. Write a receipt showing old and new dictionary versions, approval roles and deployment evidence.

Close the workshop with an independent reconstruction exercise. A reviewer who did not read the source must infer the state sequence, responsible actor, original uncertainty, version in use and permissible next action from the translation alone. Compare that reconstruction with the source record and record all differences. An apparently fluent candidate fails if it changes any of those answers. This test is useful because it measures actual decision equivalence rather than superficial word agreement.

Workshop 3: Sample to result

Build a fictional record that passes sequentially through these states: Collection requested → specimen collected → laboratory accepted → result available → report released. Translate the label at each boundary into two target-language drafts, one intended for a professional and one for a lay reader. Preserve the same event identity while changing register. A sample request does not establish that a report exists. Ask which action could be taken after each stage and which action remains premature; these are the test oracles for acceptance.

Next create an adversarial version in which the second or third state is mistakenly labelled as the final one. Trace how that claim affects a dashboard total, a participant notification, a report, a help-desk explanation and an audit trail. Where the error changes the state of evidence, mark it as semantic rather than stylistic. Finally simulate a source revision: change only one authorised condition and identify every translated asset and structured field that must be retested. Write a receipt showing old and new dictionary versions, approval roles and deployment evidence.

Close the workshop with an independent reconstruction exercise. A reviewer who did not read the source must infer the state sequence, responsible actor, original uncertainty, version in use and permissible next action from the translation alone. Compare that reconstruction with the source record and record all differences. An apparently fluent candidate fails if it changes any of those answers. This test is useful because it measures actual decision equivalence rather than superficial word agreement.

Workshop 4: Positive screening signal

Build a fictional record that passes sequentially through these states: screening signal recorded → result interpreted → patient notified → diagnostic referral submitted → diagnostic finding recorded. Translate the label at each boundary into two target-language drafts, one intended for a professional and one for a lay reader. Preserve the same event identity while changing register. A positive screen is not equivalent to confirmed disease. Ask which action could be taken after each stage and which action remains premature; these are the test oracles for acceptance.

Next create an adversarial version in which the second or third state is mistakenly labelled as the final one. Trace how that claim affects a dashboard total, a participant notification, a report, a help-desk explanation and an audit trail. Where the error changes the state of evidence, mark it as semantic rather than stylistic. Finally simulate a source revision: change only one authorised condition and identify every translated asset and structured field that must be retested. Write a receipt showing old and new dictionary versions, approval roles and deployment evidence.

Close the workshop with an independent reconstruction exercise. A reviewer who did not read the source must infer the state sequence, responsible actor, original uncertainty, version in use and permissible next action from the translation alone. Compare that reconstruction with the source record and record all differences. An apparently fluent candidate fails if it changes any of those answers. This test is useful because it measures actual decision equivalence rather than superficial word agreement.

Workshop 5: Missing data

Build a fictional record that passes sequentially through these states: question not asked → question unanswered → value unknown → valid answer recorded → record corrected. Translate the label at each boundary into two target-language drafts, one intended for a professional and one for a lay reader. Preserve the same event identity while changing register. Unknown must not collapse to no at any stage. Ask which action could be taken after each stage and which action remains premature; these are the test oracles for acceptance.

Next create an adversarial version in which the second or third state is mistakenly labelled as the final one. Trace how that claim affects a dashboard total, a participant notification, a report, a help-desk explanation and an audit trail. Where the error changes the state of evidence, mark it as semantic rather than stylistic. Finally simulate a source revision: change only one authorised condition and identify every translated asset and structured field that must be retested. Write a receipt showing old and new dictionary versions, approval roles and deployment evidence.

Close the workshop with an independent reconstruction exercise. A reviewer who did not read the source must infer the state sequence, responsible actor, original uncertainty, version in use and permissible next action from the translation alone. Compare that reconstruction with the source record and record all differences. An apparently fluent candidate fails if it changes any of those answers. This test is useful because it measures actual decision equivalence rather than superficial word agreement.

Workshop 6: Controlled terminology update

Build a fictional record that passes sequentially through these states: proposal → clinical review → linguistic review → release approval → deployment verified. Translate the label at each boundary into two target-language drafts, one intended for a professional and one for a lay reader. Preserve the same event identity while changing register. Approval in one discipline is not equivalent to every release gate. Ask which action could be taken after each stage and which action remains premature; these are the test oracles for acceptance.

Next create an adversarial version in which the second or third state is mistakenly labelled as the final one. Trace how that claim affects a dashboard total, a participant notification, a report, a help-desk explanation and an audit trail. Where the error changes the state of evidence, mark it as semantic rather than stylistic. Finally simulate a source revision: change only one authorised condition and identify every translated asset and structured field that must be retested. Write a receipt showing old and new dictionary versions, approval roles and deployment evidence.

Close the workshop with an independent reconstruction exercise. A reviewer who did not read the source must infer the state sequence, responsible actor, original uncertainty, version in use and permissible next action from the translation alone. Compare that reconstruction with the source record and record all differences. An apparently fluent candidate fails if it changes any of those answers. This test is useful because it measures actual decision equivalence rather than superficial word agreement.

Workshop 7: Multilingual fallback

Build a fictional record that passes sequentially through these states: locale requested → target string missing → source-language fallback served → user warned if needed → reviewed target string released. Translate the label at each boundary into two target-language drafts, one intended for a professional and one for a lay reader. Preserve the same event identity while changing register. The system must record which language it really served. Ask which action could be taken after each stage and which action remains premature; these are the test oracles for acceptance.

Next create an adversarial version in which the second or third state is mistakenly labelled as the final one. Trace how that claim affects a dashboard total, a participant notification, a report, a help-desk explanation and an audit trail. Where the error changes the state of evidence, mark it as semantic rather than stylistic. Finally simulate a source revision: change only one authorised condition and identify every translated asset and structured field that must be retested. Write a receipt showing old and new dictionary versions, approval roles and deployment evidence.

Close the workshop with an independent reconstruction exercise. A reviewer who did not read the source must infer the state sequence, responsible actor, original uncertainty, version in use and permissible next action from the translation alone. Compare that reconstruction with the source record and record all differences. An apparently fluent candidate fails if it changes any of those answers. This test is useful because it measures actual decision equivalence rather than superficial word agreement.

Workshop 8: Referral lifecycle

Build a fictional record that passes sequentially through these states: referral created → sent → acknowledged → accepted → follow-up completed. Translate the label at each boundary into two target-language drafts, one intended for a professional and one for a lay reader. Preserve the same event identity while changing register. No sender can claim receiver acceptance without evidence. Ask which action could be taken after each stage and which action remains premature; these are the test oracles for acceptance.

Next create an adversarial version in which the second or third state is mistakenly labelled as the final one. Trace how that claim affects a dashboard total, a participant notification, a report, a help-desk explanation and an audit trail. Where the error changes the state of evidence, mark it as semantic rather than stylistic. Finally simulate a source revision: change only one authorised condition and identify every translated asset and structured field that must be retested. Write a receipt showing old and new dictionary versions, approval roles and deployment evidence.

Close the workshop with an independent reconstruction exercise. A reviewer who did not read the source must infer the state sequence, responsible actor, original uncertainty, version in use and permissible next action from the translation alone. Compare that reconstruction with the source record and record all differences. An apparently fluent candidate fails if it changes any of those answers. This test is useful because it measures actual decision equivalence rather than superficial word agreement.

Workshop 9: Measurement reporting

Build a fictional record that passes sequentially through these states: raw result → unit recorded → method checked → decision rule applied → interpretation communicated. Translate the label at each boundary into two target-language drafts, one intended for a professional and one for a lay reader. Preserve the same event identity while changing register. Neither conversion nor interpretation may be silently introduced. Ask which action could be taken after each stage and which action remains premature; these are the test oracles for acceptance.

Next create an adversarial version in which the second or third state is mistakenly labelled as the final one. Trace how that claim affects a dashboard total, a participant notification, a report, a help-desk explanation and an audit trail. Where the error changes the state of evidence, mark it as semantic rather than stylistic. Finally simulate a source revision: change only one authorised condition and identify every translated asset and structured field that must be retested. Write a receipt showing old and new dictionary versions, approval roles and deployment evidence.

Close the workshop with an independent reconstruction exercise. A reviewer who did not read the source must infer the state sequence, responsible actor, original uncertainty, version in use and permissible next action from the translation alone. Compare that reconstruction with the source record and record all differences. An apparently fluent candidate fails if it changes any of those answers. This test is useful because it measures actual decision equivalence rather than superficial word agreement.

Workshop 10: Historical export

Build a fictional record that passes sequentially through these states: concept active in old release → old record stored → new release adopted → old code inactive → old record still interpretable. Translate the label at each boundary into two target-language drafts, one intended for a professional and one for a lay reader. Preserve the same event identity while changing register. A retired term may still be correct evidence of a historical state. Ask which action could be taken after each stage and which action remains premature; these are the test oracles for acceptance.

Next create an adversarial version in which the second or third state is mistakenly labelled as the final one. Trace how that claim affects a dashboard total, a participant notification, a report, a help-desk explanation and an audit trail. Where the error changes the state of evidence, mark it as semantic rather than stylistic. Finally simulate a source revision: change only one authorised condition and identify every translated asset and structured field that must be retested. Write a receipt showing old and new dictionary versions, approval roles and deployment evidence.

Close the workshop with an independent reconstruction exercise. A reviewer who did not read the source must infer the state sequence, responsible actor, original uncertainty, version in use and permissible next action from the translation alone. Compare that reconstruction with the source record and record all differences. An apparently fluent candidate fails if it changes any of those answers. This test is useful because it measures actual decision equivalence rather than superficial word agreement.

Workshop 11: Consent restriction

Build a fictional record that passes sequentially through these states: discussion → affirmative choice → scope documented → withdrawal requested → withdrawal handling completed. Translate the label at each boundary into two target-language drafts, one intended for a professional and one for a lay reader. Preserve the same event identity while changing register. The last two stages carry different promises. Ask which action could be taken after each stage and which action remains premature; these are the test oracles for acceptance.

Next create an adversarial version in which the second or third state is mistakenly labelled as the final one. Trace how that claim affects a dashboard total, a participant notification, a report, a help-desk explanation and an audit trail. Where the error changes the state of evidence, mark it as semantic rather than stylistic. Finally simulate a source revision: change only one authorised condition and identify every translated asset and structured field that must be retested. Write a receipt showing old and new dictionary versions, approval roles and deployment evidence.

Close the workshop with an independent reconstruction exercise. A reviewer who did not read the source must infer the state sequence, responsible actor, original uncertainty, version in use and permissible next action from the translation alone. Compare that reconstruction with the source record and record all differences. An apparently fluent candidate fails if it changes any of those answers. This test is useful because it measures actual decision equivalence rather than superficial word agreement.

Workshop 12: Public consultation to dataset

Build a fictional record that passes sequentially through these states: comment received → coded for analysis → attribution recorded → response issued → final rule effective. Translate the label at each boundary into two target-language drafts, one intended for a professional and one for a lay reader. Preserve the same event identity while changing register. An input submitted is not a policy change already in effect. Ask which action could be taken after each stage and which action remains premature; these are the test oracles for acceptance.

Next create an adversarial version in which the second or third state is mistakenly labelled as the final one. Trace how that claim affects a dashboard total, a participant notification, a report, a help-desk explanation and an audit trail. Where the error changes the state of evidence, mark it as semantic rather than stylistic. Finally simulate a source revision: change only one authorised condition and identify every translated asset and structured field that must be retested. Write a receipt showing old and new dictionary versions, approval roles and deployment evidence.

Close the workshop with an independent reconstruction exercise. A reviewer who did not read the source must infer the state sequence, responsible actor, original uncertainty, version in use and permissible next action from the translation alone. Compare that reconstruction with the source record and record all differences. An apparently fluent candidate fails if it changes any of those answers. This test is useful because it measures actual decision equivalence rather than superficial word agreement.

8. Advanced diagnostics for translators, terminology teams and reviewers

Diagnostic exercise 1: Missing vs negative

Create four sample messages: not measured, not reported, valid negative and invalid. Explain what a person can conclude from each and show why a single target word is inadequate. For marking, require a small controlled source record, two genuinely different candidate interpretations, a specific counterexample that would make the wrong interpretation visible, and a criterion by which a reviewer can accept or reject the repair. Avoid marking purely on preferred wording; a successful answer must preserve who did what, when, to whom, under which rule and with what degree of evidence.

For transfer, apply the same test to a different delivery medium and write down what changes. A screen label has space constraints; a clinical report requires traceability; a telephone script requires clear reference and an explicit handoff; an API field requires structural stability. No medium permits a different assertion merely because its style differs. The final document should expose both an example that passes and a counterexample that fails, so future reviewers can reproduce the judgment without relying on the original editor’s intuition.

Diagnostic exercise 2: Strict vs inclusive boundary

Write a miniature rule with greater than and another with greater than or equal to. Test a value exactly on the boundary in a structured record and the public explanation. For marking, require a small controlled source record, two genuinely different candidate interpretations, a specific counterexample that would make the wrong interpretation visible, and a criterion by which a reviewer can accept or reject the repair. Avoid marking purely on preferred wording; a successful answer must preserve who did what, when, to whom, under which rule and with what degree of evidence.

For transfer, apply the same test to a different delivery medium and write down what changes. A screen label has space constraints; a clinical report requires traceability; a telephone script requires clear reference and an explicit handoff; an API field requires structural stability. No medium permits a different assertion merely because its style differs. The final document should expose both an example that passes and a counterexample that fails, so future reviewers can reproduce the judgment without relying on the original editor’s intuition.

Diagnostic exercise 3: Multiple language displays

Choose a concept with an approved formal description and an approved short display. Explain how both can be correct while the underlying identifier remains unchanged. For marking, require a small controlled source record, two genuinely different candidate interpretations, a specific counterexample that would make the wrong interpretation visible, and a criterion by which a reviewer can accept or reject the repair. Avoid marking purely on preferred wording; a successful answer must preserve who did what, when, to whom, under which rule and with what degree of evidence.

For transfer, apply the same test to a different delivery medium and write down what changes. A screen label has space constraints; a clinical report requires traceability; a telephone script requires clear reference and an explicit handoff; an API field requires structural stability. No medium permits a different assertion merely because its style differs. The final document should expose both an example that passes and a counterexample that fails, so future reviewers can reproduce the judgment without relying on the original editor’s intuition.

Diagnostic exercise 4: Unsupported crosswalk

Take two plausible near-synonyms with different inclusion criteria. State why a direct mapping is unsafe and what evidence would justify an approximate relation. For marking, require a small controlled source record, two genuinely different candidate interpretations, a specific counterexample that would make the wrong interpretation visible, and a criterion by which a reviewer can accept or reject the repair. Avoid marking purely on preferred wording; a successful answer must preserve who did what, when, to whom, under which rule and with what degree of evidence.

For transfer, apply the same test to a different delivery medium and write down what changes. A screen label has space constraints; a clinical report requires traceability; a telephone script requires clear reference and an explicit handoff; an API field requires structural stability. No medium permits a different assertion merely because its style differs. The final document should expose both an example that passes and a counterexample that fails, so future reviewers can reproduce the judgment without relying on the original editor’s intuition.

Diagnostic exercise 5: Population transfer

Describe a metric with an explicit study denominator. Now move it to a different screening population and show what must remain qualified. For marking, require a small controlled source record, two genuinely different candidate interpretations, a specific counterexample that would make the wrong interpretation visible, and a criterion by which a reviewer can accept or reject the repair. Avoid marking purely on preferred wording; a successful answer must preserve who did what, when, to whom, under which rule and with what degree of evidence.

For transfer, apply the same test to a different delivery medium and write down what changes. A screen label has space constraints; a clinical report requires traceability; a telephone script requires clear reference and an explicit handoff; an API field requires structural stability. No medium permits a different assertion merely because its style differs. The final document should expose both an example that passes and a counterexample that fails, so future reviewers can reproduce the judgment without relying on the original editor’s intuition.

Diagnostic exercise 6: Consent event

Distinguish consent discussion, choice, scope of use and request to withdraw. Create a versioned notification that does not overpromise deletion. For marking, require a small controlled source record, two genuinely different candidate interpretations, a specific counterexample that would make the wrong interpretation visible, and a criterion by which a reviewer can accept or reject the repair. Avoid marking purely on preferred wording; a successful answer must preserve who did what, when, to whom, under which rule and with what degree of evidence.

For transfer, apply the same test to a different delivery medium and write down what changes. A screen label has space constraints; a clinical report requires traceability; a telephone script requires clear reference and an explicit handoff; an API field requires structural stability. No medium permits a different assertion merely because its style differs. The final document should expose both an example that passes and a counterexample that fails, so future reviewers can reproduce the judgment without relying on the original editor’s intuition.

Diagnostic exercise 7: Status chronology

Write a five-event audit trail. Translate the summary into a sentence that preserves which stage was actually completed and which was merely requested. For marking, require a small controlled source record, two genuinely different candidate interpretations, a specific counterexample that would make the wrong interpretation visible, and a criterion by which a reviewer can accept or reject the repair. Avoid marking purely on preferred wording; a successful answer must preserve who did what, when, to whom, under which rule and with what degree of evidence.

For transfer, apply the same test to a different delivery medium and write down what changes. A screen label has space constraints; a clinical report requires traceability; a telephone script requires clear reference and an explicit handoff; an API field requires structural stability. No medium permits a different assertion merely because its style differs. The final document should expose both an example that passes and a counterexample that fails, so future reviewers can reproduce the judgment without relying on the original editor’s intuition.

Diagnostic exercise 8: Historical version

Invent a code retired in a later release. Show how historical reports must preserve the older meaning while new entry is restricted. For marking, require a small controlled source record, two genuinely different candidate interpretations, a specific counterexample that would make the wrong interpretation visible, and a criterion by which a reviewer can accept or reject the repair. Avoid marking purely on preferred wording; a successful answer must preserve who did what, when, to whom, under which rule and with what degree of evidence.

For transfer, apply the same test to a different delivery medium and write down what changes. A screen label has space constraints; a clinical report requires traceability; a telephone script requires clear reference and an explicit handoff; an API field requires structural stability. No medium permits a different assertion merely because its style differs. The final document should expose both an example that passes and a counterexample that fails, so future reviewers can reproduce the judgment without relying on the original editor’s intuition.

Diagnostic exercise 9: Fallback declaration

A reviewed translation is unavailable; write a transparent fallback display. Record requested locale and served locale independently. For marking, require a small controlled source record, two genuinely different candidate interpretations, a specific counterexample that would make the wrong interpretation visible, and a criterion by which a reviewer can accept or reject the repair. Avoid marking purely on preferred wording; a successful answer must preserve who did what, when, to whom, under which rule and with what degree of evidence.

For transfer, apply the same test to a different delivery medium and write down what changes. A screen label has space constraints; a clinical report requires traceability; a telephone script requires clear reference and an explicit handoff; an API field requires structural stability. No medium permits a different assertion merely because its style differs. The final document should expose both an example that passes and a counterexample that fails, so future reviewers can reproduce the judgment without relying on the original editor’s intuition.

Diagnostic exercise 10: Approval roles

Construct a review matrix separating translator, clinical reviewer, developer, release owner and post-release monitor. Explain what each can and cannot approve. For marking, require a small controlled source record, two genuinely different candidate interpretations, a specific counterexample that would make the wrong interpretation visible, and a criterion by which a reviewer can accept or reject the repair. Avoid marking purely on preferred wording; a successful answer must preserve who did what, when, to whom, under which rule and with what degree of evidence.

For transfer, apply the same test to a different delivery medium and write down what changes. A screen label has space constraints; a clinical report requires traceability; a telephone script requires clear reference and an explicit handoff; an API field requires structural stability. No medium permits a different assertion merely because its style differs. The final document should expose both an example that passes and a counterexample that fails, so future reviewers can reproduce the judgment without relying on the original editor’s intuition.

Diagnostic exercise 11: Machine and human views

Keep the code stable while revising an accessible label. Prove that semantic identity survives a user-interface update and that the label genuinely improves comprehension. For marking, require a small controlled source record, two genuinely different candidate interpretations, a specific counterexample that would make the wrong interpretation visible, and a criterion by which a reviewer can accept or reject the repair. Avoid marking purely on preferred wording; a successful answer must preserve who did what, when, to whom, under which rule and with what degree of evidence.

For transfer, apply the same test to a different delivery medium and write down what changes. A screen label has space constraints; a clinical report requires traceability; a telephone script requires clear reference and an explicit handoff; an API field requires structural stability. No medium permits a different assertion merely because its style differs. The final document should expose both an example that passes and a counterexample that fails, so future reviewers can reproduce the judgment without relying on the original editor’s intuition.

Diagnostic exercise 12: Reconciliation

Build two services with the same code string but different system URIs. Explain the defect if a warehouse strips the URI during import. For marking, require a small controlled source record, two genuinely different candidate interpretations, a specific counterexample that would make the wrong interpretation visible, and a criterion by which a reviewer can accept or reject the repair. Avoid marking purely on preferred wording; a successful answer must preserve who did what, when, to whom, under which rule and with what degree of evidence.

For transfer, apply the same test to a different delivery medium and write down what changes. A screen label has space constraints; a clinical report requires traceability; a telephone script requires clear reference and an explicit handoff; an API field requires structural stability. No medium permits a different assertion merely because its style differs. The final document should expose both an example that passes and a counterexample that fails, so future reviewers can reproduce the judgment without relying on the original editor’s intuition.

Diagnostic exercise 13: Multi-select eligibility

Invent a value set in which two selection options overlap in ordinary language but are distinct under the rule. Give the disambiguating context that a translated label must retain. For marking, require a small controlled source record, two genuinely different candidate interpretations, a specific counterexample that would make the wrong interpretation visible, and a criterion by which a reviewer can accept or reject the repair. Avoid marking purely on preferred wording; a successful answer must preserve who did what, when, to whom, under which rule and with what degree of evidence.

For transfer, apply the same test to a different delivery medium and write down what changes. A screen label has space constraints; a clinical report requires traceability; a telephone script requires clear reference and an explicit handoff; an API field requires structural stability. No medium permits a different assertion merely because its style differs. The final document should expose both an example that passes and a counterexample that fails, so future reviewers can reproduce the judgment without relying on the original editor’s intuition.

Diagnostic exercise 14: Risk explanation

Create one population-level percentage and one individual-language caution. Distinguish uncertainty, risk and a diagnostic conclusion. For marking, require a small controlled source record, two genuinely different candidate interpretations, a specific counterexample that would make the wrong interpretation visible, and a criterion by which a reviewer can accept or reject the repair. Avoid marking purely on preferred wording; a successful answer must preserve who did what, when, to whom, under which rule and with what degree of evidence.

For transfer, apply the same test to a different delivery medium and write down what changes. A screen label has space constraints; a clinical report requires traceability; a telephone script requires clear reference and an explicit handoff; an API field requires structural stability. No medium permits a different assertion merely because its style differs. The final document should expose both an example that passes and a counterexample that fails, so future reviewers can reproduce the judgment without relying on the original editor’s intuition.

Diagnostic exercise 15: Quality-control status

Separate process-quality approval from final release of a report. Design a dashboard that cannot display released when only a QC check has passed. For marking, require a small controlled source record, two genuinely different candidate interpretations, a specific counterexample that would make the wrong interpretation visible, and a criterion by which a reviewer can accept or reject the repair. Avoid marking purely on preferred wording; a successful answer must preserve who did what, when, to whom, under which rule and with what degree of evidence.

For transfer, apply the same test to a different delivery medium and write down what changes. A screen label has space constraints; a clinical report requires traceability; a telephone script requires clear reference and an explicit handoff; an API field requires structural stability. No medium permits a different assertion merely because its style differs. The final document should expose both an example that passes and a counterexample that fails, so future reviewers can reproduce the judgment without relying on the original editor’s intuition.

Diagnostic exercise 16: Date handoff

Use publication, effective and deployed dates for one controlled term. Show how a historical export and current interface can legitimately display different authorised versions. For marking, require a small controlled source record, two genuinely different candidate interpretations, a specific counterexample that would make the wrong interpretation visible, and a criterion by which a reviewer can accept or reject the repair. Avoid marking purely on preferred wording; a successful answer must preserve who did what, when, to whom, under which rule and with what degree of evidence.

For transfer, apply the same test to a different delivery medium and write down what changes. A screen label has space constraints; a clinical report requires traceability; a telephone script requires clear reference and an explicit handoff; an API field requires structural stability. No medium permits a different assertion merely because its style differs. The final document should expose both an example that passes and a counterexample that fails, so future reviewers can reproduce the judgment without relying on the original editor’s intuition.

Diagnostic exercise 17: Source quotation

Record a person’s ambiguous free-text answer. Explain when it should remain verbatim and why forcing it into a negative code is unsupported. For marking, require a small controlled source record, two genuinely different candidate interpretations, a specific counterexample that would make the wrong interpretation visible, and a criterion by which a reviewer can accept or reject the repair. Avoid marking purely on preferred wording; a successful answer must preserve who did what, when, to whom, under which rule and with what degree of evidence.

For transfer, apply the same test to a different delivery medium and write down what changes. A screen label has space constraints; a clinical report requires traceability; a telephone script requires clear reference and an explicit handoff; an API field requires structural stability. No medium permits a different assertion merely because its style differs. The final document should expose both an example that passes and a counterexample that fails, so future reviewers can reproduce the judgment without relying on the original editor’s intuition.

Diagnostic exercise 18: Audit proof

Write an evidence packet with identifier, version, field, initial defect, corrected assertion, reviewer and production receipt. Distinguish repair of display from repair of clinical data. For marking, require a small controlled source record, two genuinely different candidate interpretations, a specific counterexample that would make the wrong interpretation visible, and a criterion by which a reviewer can accept or reject the repair. Avoid marking purely on preferred wording; a successful answer must preserve who did what, when, to whom, under which rule and with what degree of evidence.

For transfer, apply the same test to a different delivery medium and write down what changes. A screen label has space constraints; a clinical report requires traceability; a telephone script requires clear reference and an explicit handoff; an API field requires structural stability. No medium permits a different assertion merely because its style differs. The final document should expose both an example that passes and a counterexample that fails, so future reviewers can reproduce the judgment without relying on the original editor’s intuition.

Diagnostic exercise 19: Accessibility

Test a translated warning in a screen reader, a narrow display and a high-zoom view. The instruction must remain intelligible when layout changes. For marking, require a small controlled source record, two genuinely different candidate interpretations, a specific counterexample that would make the wrong interpretation visible, and a criterion by which a reviewer can accept or reject the repair. Avoid marking purely on preferred wording; a successful answer must preserve who did what, when, to whom, under which rule and with what degree of evidence.

For transfer, apply the same test to a different delivery medium and write down what changes. A screen label has space constraints; a clinical report requires traceability; a telephone script requires clear reference and an explicit handoff; an API field requires structural stability. No medium permits a different assertion merely because its style differs. The final document should expose both an example that passes and a counterexample that fails, so future reviewers can reproduce the judgment without relying on the original editor’s intuition.

Diagnostic exercise 20: Incident response

Assume two target locales show different states for the same record. Trace whether the cause is a dictionary revision, a stale cache, a mapping, a fallback or a content error. For marking, require a small controlled source record, two genuinely different candidate interpretations, a specific counterexample that would make the wrong interpretation visible, and a criterion by which a reviewer can accept or reject the repair. Avoid marking purely on preferred wording; a successful answer must preserve who did what, when, to whom, under which rule and with what degree of evidence.

For transfer, apply the same test to a different delivery medium and write down what changes. A screen label has space constraints; a clinical report requires traceability; a telephone script requires clear reference and an explicit handoff; an API field requires structural stability. No medium permits a different assertion merely because its style differs. The final document should expose both an example that passes and a counterexample that fails, so future reviewers can reproduce the judgment without relying on the original editor’s intuition.

9. Vocabulary and terminology distinctions that cannot be collapsed

Concept: The underlying clinical or workflow meaning represented by a terminology entry. A concept may have several language-specific descriptions. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Code: A symbol identifying a concept within a specified system. Never assume a short code uniquely identifies meaning without its system. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Code system: The named namespace defining valid coded concepts and their relationships. System identity matters before translation. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Display: Human-facing text associated with a coded item. Display may vary with language and audience. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Designation: A named alternative expression of a coded concept, often with language or use. Do not turn an alternate expression into a new code. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Preferred term: A term selected as preferred in a defined language or dialect context. Preferred is contextual, not universal. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Acceptable synonym: A recognised alternate expression of the same concept. Acceptable need not be preferred in every dialect. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Value set: A governed set of permitted codes for a particular purpose. Membership is not identical to a vocabulary’s complete universe. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Expansion: The evaluated contents of a value set under specified parameters and versions. A different version or date can change an expansion. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Canonical URL: Stable logical reference to a terminology artefact. Distinguish canonical identity from its retrieval address. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Version: A declared edition of a code system, value set, dictionary or translation asset. Do not substitute deployment date for content version. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Effective date: When a rule or content change starts to govern a workflow. Publication alone may not make it effective. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Release receipt: Evidence that a particular asset reached an identified environment. Approval is insufficient without delivery evidence. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Pending: A decision or process has not reached its final state. Pending is not negative. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Invalid: An observation or operation cannot be used as intended. Invalid is not a measured negative. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Unknown: A fact is not established from available evidence. Unknown is not false. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Not asked: A question was never put to the participant. Not asked differs from unanswered. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Screen-positive: A screening interpretation meeting a declared rule. Screen-positive is not confirmed diagnosis. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Predictive value: Proportion of results with a certain interpretation corresponding to condition status under stated settings. Population and prevalence context matter. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Sensitivity: Measure of detecting the target condition among those with it under defined conditions. Not the probability of disease in an individual. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Specificity: Measure of negative classification among those without the target condition under defined conditions. Not a blanket guarantee for a negative result. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Denominator: The population or observations to which a rate applies. Changing it changes the claim. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Pseudonymisation: Replacing direct identity with a pseudonym while permitting controlled linkage. Not automatically anonymity. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Free text: Narrative information not fully described by a code. Preserve nuance and provenance where appropriate. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Localisation: Adapting expression and presentation for a target-language context. Does not authorise changing the underlying clinical assertion. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Semantic equivalence: Preservation of relevant truth conditions, relations, uncertainties and actions. More demanding than grammatical resemblance. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Mapping relation: Description of how a source concept corresponds to a target terminology concept. May be approximate, broader or narrower. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Reversibility: Ability to reconstruct prior approved state and context. Needed for meaningful correction and audit. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Human review: Judgment by an appropriately competent person over defined evidence and scope. Requesting review is not completing it. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

Machine translation: Automated generation of translated text. Fluency and confidence do not constitute approval. A project glossary should link this distinction to source evidence, permitted target expressions, prohibited confusions and a sample sentence in which mistranslation would affect a real decision. Its purpose is control of meaning, not bulk accumulation of vocabulary.

10. Quality and research references: what the sources actually support

ISO 11669:2024 — Translation projects: General guidance. Provides a framework for translation project specifications, needs analysis, risk assessment, communication and workflows; it does not prescribe how to evaluate the translation output. This article applies the published high-level guidance to a practical translation workflow but does not reproduce proprietary standards text or claim official certification.

ISO 5060:2024 — Translation services: Evaluation of translation output. Addresses analytic evaluation of human, post-edited and unedited machine translation output, including error types and sampling; it does not itself govern the entire quality-assurance process. This article applies the published high-level guidance to a practical translation workflow but does not reproduce proprietary standards text or claim official certification.

ISO 17100:2015 — Translation services requirements. Addresses requirements for core translation processes and resources. The current published 2015 edition has a newer edition in development; its scope does not include raw machine translation plus post-editing. This article applies the published high-level guidance to a practical translation workflow but does not reproduce proprietary standards text or claim official certification.

HL7 FHIR R4 datatype definitions. Distinguishes coding references and human-language concept text, supporting the need to separate code identity from labels and user-recorded wording. This article applies the published high-level guidance to a practical translation workflow but does not reproduce proprietary standards text or claim official certification.

HL7 FHIR R4 ValueSet guidance. Describes canonical identification, versions, expansions and language/display-related parameters relevant to reproducible controlled lists. This article applies the published high-level guidance to a practical translation workflow but does not reproduce proprietary standards text or claim official certification.

SNOMED CT preferred-term guidance. Explains preferred terms as language- and dialect-dependent descriptions of concepts, not new concepts. This article applies the published high-level guidance to a practical translation workflow but does not reproduce proprietary standards text or claim official certification.

W3C language-tag guidance. Explains use of language tags in HTML and XML, which matters when a multilingual label needs correct language metadata. This article applies the published high-level guidance to a practical translation workflow but does not reproduce proprietary standards text or claim official certification.

WHO ethics and governance of artificial intelligence for health. Provides wider principles for autonomy, accountability, equity and transparency when health-related technology is used; it is not a code-translation specification. This article applies the published high-level guidance to a practical translation workflow but does not reproduce proprietary standards text or claim official certification.

Reference check: 19 November 2026. The article uses publicly accessible official abstracts and specification guidance rather than asserting access to the full paid text of an ISO standard. FHIR implementations should check the version and implementation guide actually governing their deployment; examples here are illustrative rather than production-ready coding instructions.

11. Transfer to other domains: a controlled analogy, not an imported answer

Education records: A graded assessment result is different from an incomplete attempt. The transferable method is to separate the proposition, the identifier, the permitted states, the dated authority and the evidence of completion before translating any display text. Do not infer school policy from a medical screening model. Applying the same analysis across sectors teaches the habit of semantic control without pretending that medical terminology, legal processes and school records share identical rules.

Public benefits: Application submitted is distinct from benefit approved. The transferable method is to separate the proposition, the identifier, the permitted states, the dated authority and the evidence of completion before translating any display text. Eligibility criteria and legal authority must come from the applicable scheme. Applying the same analysis across sectors teaches the habit of semantic control without pretending that medical terminology, legal processes and school records share identical rules.

Product safety: A warning issued is distinct from a recall completed. The transferable method is to separate the proposition, the identifier, the permitted states, the dated authority and the evidence of completion before translating any display text. Follow the relevant safety authority and product-specific records. Applying the same analysis across sectors teaches the habit of semantic control without pretending that medical terminology, legal processes and school records share identical rules.

Financial reporting: Transaction pending is distinct from settled. The transferable method is to separate the proposition, the identifier, the permitted states, the dated authority and the evidence of completion before translating any display text. Accounting categories require their own controlled terminology. Applying the same analysis across sectors teaches the habit of semantic control without pretending that medical terminology, legal processes and school records share identical rules.

Research datasets: A dataset field label is distinct from its typed variable definition. The transferable method is to separate the proposition, the identifier, the permitted states, the dated authority and the evidence of completion before translating any display text. Study protocols and participant permissions take precedence. Applying the same analysis across sectors teaches the habit of semantic control without pretending that medical terminology, legal processes and school records share identical rules.

Public consultation: A submitted proposal is not an adopted rule. The transferable method is to separate the proposition, the identifier, the permitted states, the dated authority and the evidence of completion before translating any display text. Decision authority and effective date must be established separately. Applying the same analysis across sectors teaches the habit of semantic control without pretending that medical terminology, legal processes and school records share identical rules.

Transport logistics: A shipment sent is not a shipment received. The transferable method is to separate the proposition, the identifier, the permitted states, the dated authority and the evidence of completion before translating any display text. Acknowledgement requires evidence at the receiver. Applying the same analysis across sectors teaches the habit of semantic control without pretending that medical terminology, legal processes and school records share identical rules.

Digital services: A password reset requested is not a credential already changed. The transferable method is to separate the proposition, the identifier, the permitted states, the dated authority and the evidence of completion before translating any display text. Authentication status must not be inferred from notification state. Applying the same analysis across sectors teaches the habit of semantic control without pretending that medical terminology, legal processes and school records share identical rules.

12. FAQ: practical decisions before final approval

1. Should a translator translate code identifiers?

Normally no: preserve the code and namespace as data while translating approved human-facing labels. If a cross-code-system mapping is required, it is a separately authorised semantic mapping, not a linguistic substitution. A complete answer also names the evidence that would reverse the conclusion and the reviewer permitted to approve that reversal. Mere familiarity of a phrase is not a substitute for the source field definition.

2. Can a glossary replace a data dictionary?

No. A glossary names preferred expressions; a data dictionary defines field meaning, type, permitted states, scope, relationships, null policy and provenance. A complete answer also names the evidence that would reverse the conclusion and the reviewer permitted to approve that reversal. Mere familiarity of a phrase is not a substitute for the source field definition.

3. Can two displays differ and still be faithful?

Yes, if they refer to the same governed concept and preserve the needed decision in their intended contexts. Variation in register is not permission to alter a workflow state. A complete answer also names the evidence that would reverse the conclusion and the reviewer permitted to approve that reversal. Mere familiarity of a phrase is not a substitute for the source field definition.

4. What if the source is ambiguous?

Do not create certainty by guessing. Record the ambiguity, request an accountable clarification and keep the affected release gate pending. A complete answer also names the evidence that would reverse the conclusion and the reviewer permitted to approve that reversal. Mere familiarity of a phrase is not a substitute for the source field definition.

5. Can an unknown value be shown as no?

Only where the source’s own explicit definition genuinely supports that mapping, which should be demonstrated. The default safe assumption is that unknown and negative are distinct. A complete answer also names the evidence that would reverse the conclusion and the reviewer permitted to approve that reversal. Mere familiarity of a phrase is not a substitute for the source field definition.

6. Does a screen-positive result confirm disease?

No. Screening result terminology must be understood under the named screening programme and should not be rewritten as a confirmed diagnosis without evidence. A complete answer also names the evidence that would reverse the conclusion and the reviewer permitted to approve that reversal. Mere familiarity of a phrase is not a substitute for the source field definition.

7. Is a translated value-set display the same as a new code?

No. The display is human-oriented; the code and its system identify the coded concept. A mapping between systems requires separate justification. A complete answer also names the evidence that would reverse the conclusion and the reviewer permitted to approve that reversal. Mere familiarity of a phrase is not a substitute for the source field definition.

8. How do I test a clinical threshold translation?

Protect the comparison operator, value, unit, method and interpretation context. Include a boundary example exactly equal to the threshold. A complete answer also names the evidence that would reverse the conclusion and the reviewer permitted to approve that reversal. Mere familiarity of a phrase is not a substitute for the source field definition.

9. What if a new term becomes preferred?

Update the approved displays for their applicable language or dialect and release scope; preserve the interpretation of historical records under their original version. A complete answer also names the evidence that would reverse the conclusion and the reviewer permitted to approve that reversal. Mere familiarity of a phrase is not a substitute for the source field definition.

10. Do ISO 11669 and ISO 5060 do the same job?

No. The former concerns translation project planning and specification; the latter concerns evaluation of translation output. A complete answer also names the evidence that would reverse the conclusion and the reviewer permitted to approve that reversal. Mere familiarity of a phrase is not a substitute for the source field definition.

11. Does professional review prove clinical validity?

No. A translation reviewer evaluates expression and fidelity within competence; clinical validity and code-system governance require designated domain authority. A complete answer also names the evidence that would reverse the conclusion and the reviewer permitted to approve that reversal. Mere familiarity of a phrase is not a substitute for the source field definition.

12. What should a version manifest contain?

At minimum source dictionary revision, code-system and value-set references, language tags, translation asset versions, sign-offs, effective dates, deployment evidence and rollback targets. A complete answer also names the evidence that would reverse the conclusion and the reviewer permitted to approve that reversal. Mere familiarity of a phrase is not a substitute for the source field definition.

13. What if the receiving system changes a display?

Compare the stored code and meaning separately from the rendered label, identify which layer caused the discrepancy and correct that owner without silently changing clinical facts. A complete answer also names the evidence that would reverse the conclusion and the reviewer permitted to approve that reversal. Mere familiarity of a phrase is not a substitute for the source field definition.

14. When should I stop the release?

Stop when a mistranslation changes a permitted state, classification boundary, identity, unit, action, consent scope, source authority, release version or user safety interpretation. A complete answer also names the evidence that would reverse the conclusion and the reviewer permitted to approve that reversal. Mere familiarity of a phrase is not a substitute for the source field definition.

15. What does complete post-release verification look like?

Inspect the public-facing output in each affected locale, test the structured message and record the results against the approved source version and acceptance criteria. A complete answer also names the evidence that would reverse the conclusion and the reviewer permitted to approve that reversal. Mere familiarity of a phrase is not a substitute for the source field definition.

13. Next routes without changing protected content owners

Return to Master Art of Translation | The Complete System for Moving Meaning Between Languages for the complete meaning-transfer framework, project design, source analysis, equivalence, terminology and human revision. The previous reserved-tens specialist child is the Screening Public Consultation, Stakeholder Submission, Dissent Record and Policy-Response Fidelity System; it covers deliberative records and policy responses, not the clinical data-dictionary and code-system layer owned here. The reserved child number of this new article is EDKSG-TRANS-MASTER-WORLD-3300; it adds a new specialist route rather than rewriting the root or an earlier article.

When the question is about learning word meanings, synonyms and durable lexical knowledge, use the protected Vocabulary Learning Hub. When the question is about English grammar, clause structure, paraphrase, attribution and how English itself expresses states, use the protected How English Works | The Complete Map and Learning Routes. Those are neighbouring learning owners, not material to copy, retitle or replace. They are linked only where the reader’s problem genuinely crosses the boundary.

The final test is concrete. Can an independent reader of either language say which event occurred, which clinical or workflow conclusion is warranted, which authority defined the code, which version applies, what remains unknown, and what action is actually permitted next? Can an independent system reconstruct the same code identity and selection constraints? If both tests pass with recorded evidence, the translation has preserved meaning across not only languages but also the structured systems that carry it.