Master Art of Translation | The Multilingual Translation Error Reporting and Correction System — How User Feedback, Severity, Reproduction, Evidence and Revalidation Preserve Meaning

How to report a bad translation, correct it safely and prove the correction worked

How do you report a translation error on a live website, multilingual app, public document or customer service channel—and ensure that the corrected translation actually reaches the people who need it? A reliable translation-error reporting system must capture the exact faulty text, language, context, source version, user impact and reproducible evidence; then it must assign a correction, verify the new wording in the actual product, and keep an auditable record of the outcome. That is more than proofreading. It is the discipline of detecting and repairing meaning after publication.

People searching for how to fix a wrong translation, report a localization bug, handle translation quality feedback or resolve multilingual content defects often find a generic support form or an instruction to contact a translator. The hidden problem is that a comment such as ‘this sentence is wrong’ is rarely enough to reproduce the failure. The apparent error may come from an outdated source, ambiguous terminology, a hidden user state, a locale fallback, automated translation, a screen-reader label, a stale cache or a correction deployed to the wrong surface. Correct words in one editing system do not prove correct words reached the reader.

This long-form guide teaches a repeatable, world-facing method for translation feedback and correction: collect a useful report, protect the reporter, reproduce the displayed text, recover the source meaning, rank consequence rather than embarrassment, select an authorised change, verify all affected variants and close the loop with evidence. It includes diagnostic cases, counterexamples, test designs and escalation routes for human and AI-assisted translation. It is one specialist child of the Master Art of Translation root; vocabulary acquisition belongs to the protected Vocabulary Learning Hub, while English-only grammar and writing belong to How English Works.

Architecture record: EDKSG-TRANS-MASTER-WORLD-1210. Scope: post-publication translation errors and their verifiable correction across languages and media. This chapter does not replace the general translation quality specification chapter, the semantic change-control chapter, the pre-release regression testing chapter, or the multilingual web accessibility chapter. It addresses the different operational moment when an actual reader, operator, reviewer or system raises a problem after content is in use.

A translation correction is complete only when the right meaning has been restored in the right language, on the right version, for the right audience—and someone has checked the released result.

1. The search intent: what does a reader actually need when a translation seems wrong?

An end user normally wants a fast answer: the wording is confusing, an instruction seems unsafe, an amount seems inconsistent, or the translation contradicts what the interface does. They do not want a taxonomy of translation errors before being heard. Put an immediate reporting route near the content, allow a screenshot or quotation, let the person select the affected language, and tell them what will happen next. A reporter should not have to invent a technical bug description. Good intake design turns lived confusion into a question that a responsible team can investigate.

An editor needs a different answer. The editor must know which source version generated the target, whether the string is reusable, which locales and channels inherit it, whether product behaviour or only text is wrong, and whether updating one field will change historical records. A project manager needs priority, an owner and an acceptance test. A translator needs sufficient context to distinguish an error from a justified variant. A user-support representative needs safe interim wording without promising an unapproved fix. Different roles need distinct information, but one case identifier should connect their work.

Search phrases such as bad translation, mistranslation correction, wrong language in app, localization bug report, incorrect subtitle, translation feedback form and multilingual QA express different stages of a common need. Build a system that starts with the plain reader question and expands only when necessary. The correct public-facing page helps readers act. The internal diagnostic record helps experts investigate. Confusing those layers makes both experiences worse.

2. The hidden problem: a sentence can be corrected yet the user still receives the error

Content rarely travels directly from translator to reader. It may move through a source repository, translation management system, content-management system, application package, API, cache, browser, notification service, PDF export, mobile release or assistive-technology layer. Each transfer can create an independent version. Editing a translation in a termbase may change no visible text until a release is rebuilt. Correcting a web page may leave the same error inside an email template or a saved offline copy. A public notice can remain dangerous even after the newest page has been fixed if the old version keeps circulating.

A complete report therefore describes two separate facts: what the intended message should mean and what readers demonstrably see or hear. A translation reviewer may find the target grammatically acceptable while the interface pairs it with the wrong control. A language variant may be correct when opened directly but incorrect after a fallback. A machine-translation preview may be corrected while an automatic retranslation overwrites it at the next source change. Each failure requires tracing the delivery path, not merely replacing words.

The practical question changes from ‘Have we fixed the text?’ to ‘Which released artefacts still carry the defective meaning?’ That question demands version identifiers, language and region settings when relevant, distribution channels, timestamps, cached variants and a test that reproduces the consumer experience. The outcome is not a more bureaucratic process; it is a shorter route from evidence to a working correction.

3. Definitions that stop teams talking past one another

A translation error changes, obscures, omits or misrepresents relevant source meaning in the target. A localisation defect can also concern formats, interactions, culturally inappropriate content, layout or functional behaviour. An internationalisation defect is a design or implementation limitation that makes correct multilingual behaviour difficult or impossible. A source-content defect is a problem already present in the original and must be corrected at its source owner. A reviewer disagreement is not automatically an error; two legitimate alternatives may differ in style, politeness or regional usage.

A report is the claim that something may be wrong. A confirmed issue is a reproducible discrepancy supported by evidence. A correction proposal is a candidate solution, not proof of release. A resolved ticket records that the action chosen by the team has been completed, while a verified correction records evidence that the intended meaning is now correctly delivered. Closure is a communication and governance decision: it should explain what changed, what did not, and whether the original reporter can meaningfully confirm the improvement.

Some defects are semantic, such as reversing permission or negation. Others are pragmatic, such as turning a gentle recommendation into an obligation. Some are technical, such as a truncated target string; others are social, such as a phrase that is insulting in an unintended context. Make the categories descriptive rather than punitive. They are tools for choosing the right reviewer and check, not labels for blaming the person who originally translated a sentence.

4. The basic invariant: separate the meaning contract from the word string

Start each investigation with a tiny meaning contract. Write down the actor, action, object, time, condition, exception, degree of certainty and intended reader response. For the sentence ‘You can cancel without a fee until the day before delivery’, the contract contains permission to cancel, no cancellation fee, and a deadline tied to delivery. A target phrase might be elegant but still fail if it changes ‘until’ to ‘after’ or introduces an unconditional cancellation right. The contract makes that failure visible even to reviewers who prefer different stylistic choices.

Not every element matters equally in every medium. Poetry can place unusual weight on rhythm, ambiguity and emotional image. A user interface may prioritise control identity and next action. A contract may require exceptional precision about scope, conditions and duties. The point is not to impose one translation theory on all content. The point is to name the invariants for this text before a reviewer proposes a fix. That avoids the common situation in which the most fluent option wins because nobody has recorded the underlying obligation.

Keep a separate change-permission statement. It can say that punctuation may change, sentence order may change, terminology may be normalised to an approved glossary, but facts, quantities, conditions and warnings cannot be invented or suppressed. This becomes an acceptance criterion. A good correction often uses different words from the source. Its legitimacy comes from preserved meaning and function rather than surface resemblance.

5. A usable report form: collect the minimum evidence without deterring people

A good report form asks for the language the person was using, where they saw the text, an exact quotation or screen capture, what confused them, and what they expected to understand or do. The form can collect an optional device or application version when the problem is technical. It should never require a reporter to identify the source language correctly, provide their legal name, or submit sensitive documents simply to flag a confusing sentence. A free-text route is essential for people who cannot describe the error in specialist terms.

Internally, attach a report identifier, observation time, content or resource identifier, target locale, source revision when known, release channel, severity hypothesis, reporter permission for follow-up and retention policy. Store the original reported wording before anyone edits it. Preserve screenshots only when useful and authorised; redact account names, addresses, payment details and medical information. If a screenshot exposes private information, record the linguistic evidence in a safer excerpt and restrict access to the original.

Design the form in the reporter’s language when possible. ‘Report a translation problem’ is clearer than ‘submit an i18n incident’. If automated language identification is used, offer a way to correct it. A contact channel that is itself untranslated can exclude exactly the readers the team needs to hear from. Confirm receipt without suggesting that correctness has already been established or promising a timetable the team cannot meet.

6. The evidence bundle: how to make a vague complaint reproducible

A reproduction bundle has five anchors: content identity, displayed wording, context, system state and a reliable observation. Content identity might be a URL, resource key, document version, notification type or timestamped subtitle cue. Displayed wording is the exact source and target string when available. Context is the adjacent instruction, user journey or prior exchange. State includes the chosen language, product version, account role, permissions and any important selection. Observation explains what the reporter actually saw or heard and what they tried next.

Do not treat a screenshot as a complete explanation. It can omit tooltips, hidden labels, earlier steps, source files and fallback logic. Do not treat a machine-generated translation as the canonical source unless the organisation has explicitly assigned it that role. Ask whether the problem appears on another device, for a different user, in an offline export, or only after changing languages mid-session. These comparisons distinguish a bad translation from a stale cache, missing string, broken encoding, mixed-language component or a genuine product defect.

A useful reproduction statement is short enough to run: open the saved test account, select the specified language, navigate to the named state, trigger the error, compare the rendered notice with the approved source and record the result. Preserve a before image and a text-based transcript where possible. Add both the human interpretation and the technically observed result, so later reviewers can understand the problem even if the original screen no longer exists.

7. Severity is not the same as linguistic ugliness

A typo in a decorative heading can be embarrassing yet low impact. An almost invisible reversal of ‘do not’ can alter a safety instruction. Evaluate consequence, exposure and uncertainty separately. Consequence asks what a reasonable reader might do because of the wording. Exposure asks how widely the defective variant is distributed and whether vulnerable or dependent readers are likely to rely on it. Uncertainty asks whether the source and intended action are verified; low confidence can justify a temporary hold rather than a confident rewrite.

Use a practical severity ladder: critical when a mistranslation plausibly causes immediate harm, denied rights or irreversible action; high when it misstates eligibility, consent, money, essential procedure or a significant claim; medium when it blocks a task or creates material misunderstanding; low when function survives but style, fluency or presentation needs improvement. These are editorial triage examples, not ISO-defined severity levels. Organisations should adapt them to their subject matter and responsibility.

Prioritisation is also temporal. A low-impact string appearing to millions of readers may warrant an early change; a high-impact notice seen by only a few may require immediate containment. Record who made the priority decision and the rationale. Never let raw user-vote counts alone decide whether a genuine safety or rights issue matters.

8. Reproduction and counterexamples: how to know what kind of defect occurred

Reproduce the reported path before changing anything. Then try controlled counterexamples. Switch only the locale while keeping account state the same. Change the account state while keeping the locale fixed. Compare the web page with the app. Compare the current source revision with the last approved one. Disable translation memory suggestions or automatic rewriting in a safe test environment. Each comparison isolates one variable and often reveals whether the defect was linguistic, technical or historical.

If the source says ‘not eligible’ and the target says ‘eligible’, the fault is semantic. If every target string is correct but appears beneath the wrong button, it is a mapping or interface defect. If the target is correct in the editor yet wrong in production, it is a release or cache defect. If the source itself has two conflicting instructions, it is a source governance defect. If a reviewer simply prefers a more idiomatic synonym, the case may be a style decision rather than a reportable failure.

Document negative evidence too. ‘Could not reproduce in a fresh session’ is useful but does not make the reporter’s observation imaginary. Record the tested conditions and seek the missing state. When a report cannot be reproduced and no risk signal remains, explain that limitation honestly and keep a reopening path. Closing a ticket because one engineer cannot reproduce a multilingual context is a weak conclusion.

9. The correction decision: who can authorise which change?

Do not allow every reviewer to change every class of statement. A language professional can normally propose natural target wording and test equivalence. A product owner controls intended feature behaviour and interface labels. A policy owner controls contractual, eligibility or consent language. A subject-matter expert must confirm specialised technical claims. A safety owner can require temporary suspension of a dangerous message. The person who approves the source meaning need not be the person who writes the target sentence.

Record the decision as an explicit mapping: old target, proposed target, source revision, affected locale, semantic difference, approver and verification plan. If the source meaning is unclear, return to the owner instead of asking a translator to guess. A good editor may supply two plausible target options; the approval record should state which condition makes one correct. That is especially important when nuance concerns obligation, uncertainty, honorifics, gendered reference or legal scope.

A correction can involve a text patch, source amendment, schema change, string-identifier change, product behaviour change, training update, termbase revision or temporary withdrawal. Choose the smallest safe solution, not the smallest visible edit. A one-word patch is not safe if the old meaning persists in notifications, cached files and print-ready documents.

10. Deployment verification: editing is not publishing and publishing is not delivery

After an approved correction, identify all output artefacts. A web content system may render server pages, search snippets, email alerts and structured data. An app may ship different translation bundles for different versions. A PDF may be mirrored or downloaded. A subtitle file may be copied into several video renditions. Test the exact target surface and version that readers use; do not substitute a successful editor preview for release evidence.

Build a small release matrix: each affected locale against its relevant channels, states, content versions and assistive-technology pathways. Include default language, user-selected language and fallback behaviour. Test cold cache and warm cache where relevant. For a message involving a number, condition, exception or reversible action, compare the released target against the approved meaning contract. If a source update is pending, confirm that it will not overwrite the correction.

A robust record states what was verified and what remains outside the test. ‘Passed on browser A for language X’ is not proof of universal correctness. If a secondary channel cannot be tested, disclose the gap and decide whether to block, monitor or release with controls. Confidence comes from named evidence, not from a completion checkbox.

11. Closure and public trust: how to acknowledge a correction without exposing private details

The person who reported the issue should receive a meaningful outcome where a contact channel exists and permission allows it. A good closure note explains the problem in plain language, says whether wording, source content or product behaviour changed, and gives an appropriate way to recheck it. It does not expose translators’ names, internal credentials, private conversations or unrelated user records. It also does not imply that one report proves all languages are now perfect.

Corrections that affect essential instructions may require a more visible notice or a targeted re-contact of affected recipients, subject to the service’s lawful permissions and policies. If a prior version caused a material misunderstanding, merely updating a hidden string may be inadequate. Keep the old and new wording in restricted issue records as needed for accountability, but avoid republishing harmful or private material just to demonstrate transparency.

Sometimes the honest outcome is ‘translation accurate; source policy unclear’ or ‘cannot confirm because the release was superseded’. A responsible system can close with a documented limit, request more evidence, or escalate to a different owner. Trust grows when a team distinguishes confirmation, uncertainty and next steps rather than presenting every ticket as solved.

12. The relationship between translation evaluation and correction

ISO 5060:2024 offers guidance for evaluating translation output using analytic error types and penalty approaches, including human, post-edited and unedited machine output. Its published scope explicitly excludes the procedures for assuring quality and corrective actions. That distinction matters: scoring a target text does not by itself tell a service how to locate a released defect, contain its effects, notify readers and verify a fix. This chapter builds a practical correction workflow beside evaluation, without pretending to quote hidden or proprietary clauses of the standard.

ISO 11669:2024 frames translation projects through specifications, needs analysis, risks and communication among parties. ISO 17100:2015 addresses the processes and resources necessary for defined translation services, while its stated scope does not include raw machine output plus post-editing. ISO 18587:2017 addresses full human post-editing of machine translation output and was still listed as published with a revision project under way at the time of this research. The appropriate lesson is to read each standard for its actual remit, not to use its number as a generic quality endorsement.

Evaluation is a snapshot of a defined output under a defined rubric. Incident handling is a process triggered by evidence that a specific released variant may have failed. Both need competent human judgement, but they answer different questions. Link an evaluation score to a reproducible defect only where there is a factual relationship. A single high-impact error can warrant action even if a sampled quality score looks acceptable.

13. Machine translation and generative AI: a fluent correction can still be wrong

Automated systems can accelerate detection, propose alternative wording, compare numbers and identify inconsistent terms. They can also fabricate explanations, remove ambiguity, insert unwarranted certainty or confidently claim that a preferred phrasing is universally correct. Use an AI suggestion as a candidate supported by evidence, not as a finding of fact. Keep the original source, prompt or configuration version when relevant, proposed target and human acceptance reason. Never paste unnecessary personal data into an external translation service as a shortcut.

Where a defect arises from automated translation, ask whether it is isolated, systematic or triggered by context. A model might repeatedly erase negation in short fragments, mistranslate one polysemous term, mishandle placeholder order or transform labels after a source update. Correcting one string without fixing the generation rule may produce the same failure tomorrow. Determine which cases require a controlled glossary, human review queue, additional source context, restricted automation or a different publishing boundary.

NIST’s AI Risk Management Framework and its 2024 generative AI profile provide voluntary risk-management context for designing responsible oversight; they are not translation-specific certification. Use them as prompts to identify roles, review risk and track observed failures, not as evidence that an AI product is trustworthy merely because it uses the vocabulary of governance.

14. Language metadata and accessibility during a correction

A displayed sentence can be linguistically right but inaccessible when the page marks the wrong language. Language tags influence assistive reading and pronunciation. W3C guidance on WCAG 2.2 criterion 3.1.2 addresses identifying the human language of passages or phrases, with defined exceptions. When investigating reports from screen-reader users, capture both the visible text and the programmatic language context. Fixing spelling without correcting metadata may leave the functional failure unchanged.

An error-report form must itself be understandable. W3C criterion 3.3.1 requires text identification and description of automatically detected input errors. That principle has a direct operational implication: readers who struggle to submit a localization report need to know what failed and how to continue. Localised form labels, status notices, keyboard access and feedback after submission are part of the reporter’s route, not optional decoration.

Do not copy the entire accessibility programme into this article. The protected earlier chapter owns comprehensive multilingual accessible-name, form and screen-reader localisation. Here the focus is narrower: the evidence necessary to reproduce an accessibility-linked translation report, the owner who can approve the fix, and the proof that the repaired language is actually exposed through the relevant channel.

15. The correction ledger as an evidence system

Use a correction ledger to separate observations, hypotheses, decisions and verification. Minimum fields include incident identifier, reported text, source reference, target locale, affected surfaces, status, reviewer, severity rationale, corrective action, test result and closure. Make state changes explicit: received, triaged, awaiting source clarification, confirmed, corrected, deployed, verified, closed or reopened. Those labels are organisational choices rather than a universal mandated workflow.

Do not preserve every raw screenshot forever. Choose retention proportionate to risk and obligations, and restrict private evidence. A screenshot from a user’s account may contain unrelated details; a text excerpt and resource key may be sufficient after the initial investigation. A public article can teach the process without publishing any actual private correction cases. Every case study below uses invented illustrative scenarios and English paraphrases rather than claiming access to real incident databases.

A change history should allow a reviewer to answer four questions months later: what readers could have seen, why it was considered wrong, who accepted the correction and what proof showed the released fix. It should also reveal whether a later source revision reverted the change. That is a useful design target whether the team uses a spreadsheet, issue tracker, translation platform or integrated content pipeline.

16. Metrics that encourage repair rather than silence

Count reports received, valid reports, high-consequence failures, recurring defect classes, time to usable acknowledgement, time to verified fix, and reopened cases. Break results down by target language, channel and content family only when the sample is meaningful. Do not rank languages by raw complaint count without considering reader population, reporting access and differences in risk. A language receiving few reports may have a bad feedback route rather than excellent translations.

Record correction escape rate: how often a supposedly fixed issue appears again in another channel or release. Track missing evidence rates to see whether forms and identifiers are usable. Separate linguistic disputes from confirmed semantic errors to prevent style preferences overwhelming risk prioritisation. Pair quantitative metrics with a handful of carefully investigated narratives, because a single severe mistranslation can carry more learning value than a thousand low-impact spelling tickets.

Do not reward a team for closing tickets quickly while leaving readers confused. Make verified closure more important than raw closure. A well-designed metric rewards discovery of hidden errors rather than penalising the people who reveal them. Users should feel safe to report faults; reviewers should feel safe to record uncertainty; owners should have a clear path to correct their systems.

17. A repeatable twelve-step method from report to release evidence

Step one is acknowledgement in the reporter’s language where feasible. Step two captures the original display without making public claims about fault. Step three records the relevant language, content identifier and state. Step four creates a meaning contract. Step five assesses plausible harm and exposure. Step six reproduces and isolates the defect using controlled comparisons. If the meaning contract cannot be determined, the process pauses for a source-owner decision rather than manufacturing a fluent guess.

Step seven selects and authorises the change, including whether source content, code, terminology or target language must change. Step eight identifies all dependent channels. Step nine prepares the target correction and checks it against the source and user action. Step ten publishes or deploys according to the normal change process. Step eleven tests the exact released variant and affected fallbacks. Step twelve closes or reopens with a record of evidence and a response that the original reporter can understand.

The sequence is adaptable. A dangerous false instruction may justify temporary containment before a complete root-cause analysis; a trivial punctuation issue may pass through a shorter queue. Adapt effort to consequence without discarding the logical stages. A good system can be light for ordinary content and rigorous for safety-critical communication.

18. Four failure modes that masquerade as successful corrections

First is the editor-only fix: a field looks right in a translation platform but an application bundle still contains the old string. Second is the one-locale fix: the reported language is repaired but an inherited regional variant and the generic fallback remain wrong. Third is the copy-only fix: the visible text is repaired but the associated control, state or link is still misleading. Fourth is the temporary fix: a nightly regeneration process rebuilds the bad target from unchanged instructions and silently undoes the patch.

Each failure needs a different diagnostic test. For editor-only defects, compare deployed build hashes or version identifiers where available. For locale inheritance, enumerate actual fallback paths and test them. For copy-only defects, observe the user journey and its consequences rather than merely proofreading. For regeneration defects, repeat the process that originally produced the target and confirm the error does not return. A screenshot taken once, immediately after release, will miss some of these patterns.

These are not exotic edge cases. They arise whenever language crosses organisational and technical boundaries. The cure is simple in principle and demanding in practice: keep a traceable source-to-delivery chain and make release verification part of the correction, not an optional extra.

19. Correction ethics: respect the reporter, protect context and avoid invented consensus

Language identity and preference are socially meaningful. When users report an expression as rude, exclusionary or unnatural, listen carefully without treating one individual as the spokesperson for every user of the language. Consider register, community, audience and evidence. Localised wording can be appropriate for one audience and inappropriate for another. An authorised style decision should record which audience it serves; it should not erase legitimate disagreement under the label of objective grammar.

A report involving identity, disability, trauma or confidential communication needs sensitive handling. Collect only the linguistic facts required for diagnosis. Do not ask someone to disclose a diagnosis to prove that a screen-reader label is unhelpful. Do not force someone to provide a legal document when a short anonymised excerpt is sufficient. Explain when specialists or affected community reviewers will be consulted, and distinguish consultation from endorsement.

Corrections sometimes require saying the previous text was wrong. That can be done plainly without humiliation. The objective is restoring meaningful communication, not identifying a scapegoat. For high-stakes communications, a second independent reviewer and an explicit source-owner check may be more valuable than a public apology that lacks a reliable remedial route.

20. Transfer: how the same method works outside software

A book erratum, a museum label, a printed warning, an online article, an audiobook, a caption track and an appointment notice have different distribution systems. All can use the same logic: preserve the reported instance, locate the authoritative source, identify the target audience and version, decide whether meaning failed, approve the remedy, distribute the corrected artefact and provide a path for people still holding the older version.

Where text cannot be retracted, the correction strategy may be an errata page, a replacement insert, updated metadata, a notice at the point of access or a direct clarification to affected readers. The right choice depends on scope, law, risk and practical reach. Do not claim that a digital edit changes printed copies or that a revised subtitle automatically repairs a clip distributed elsewhere. Explicitly document the distribution limit.

The method even applies to informal community translation. A volunteer who sees a confusing notice can report the exact phrase and describe the likely misunderstanding. A coordinator can compare the source, check local usage, approve a correction and notify the community. Professional tools help at scale, but disciplined attention to meaning is the fundamental resource.

21. How to interpret a disagreement between two competent translators

Disagreement may be evidence of a defect, but it may also indicate several legitimate strategies. Ask each reviewer to explain which source feature they are protecting and which audience requirement they are applying. One version may be more literal, another more idiomatic. The decisive question is whether either version misstates a required fact, changes an action, violates the agreed style constraints or loses an intended implication. Do not choose by vote when the real question concerns a source ambiguity only its owner can resolve.

If the dispute turns on terminology, consult the authorised glossary and authentic target-language examples from the same domain. If it concerns politeness, provide the relationship and communicative setting. If it concerns a regulated label, obtain approved reference material. If both are acceptable, record the preferred style as a future convention but avoid classifying a previous correct version as a serious mistranslation simply because the wording changed.

Teach reviewers to articulate counterexamples. If version A makes the instruction apply to all users but version B limits it to verified users, that is not mere style: it changes scope. If version A uses a formal verb and version B a neutral one while both preserve rights, the difference may be a legitimate editorial choice. Evidence gives the team a way out of taste-based argument.

22. Templates for triage, correction decisions and reader responses

A good triage note can be written as a compact paragraph: ‘In target language X on surface Y, the released message says Z when the approved source means W. The issue is reproducible in state S. The likely consequence is C. We propose action A, require owner O and will accept the change only after test T passes.’ That is more useful than a lengthy argument over whether a phrase sounds poor.

A decision record can say: ‘We considered literal and functional variants. We selected the latter because it preserves the actor, condition and action while using normal target-language syntax. The source owner approved the deadline and the language reviewer approved the target expression. We will verify the web, email and downloadable document variants.’ If agreement is missing, mark the uncertainty openly rather than quietly publishing.

A reader response can say: ‘Thank you for identifying the confusing wording. We confirmed the issue and corrected the instruction in the current version. The earlier version could be understood as allowing an action after the deadline; the corrected wording states the deadline clearly. Please reopen this report if you still see the older wording.’ Use a different response when the case was not reproduced or when the source itself remains under review. Do not promise universal completeness.

Worked diagnostic laboratory: 100 specific reports and how to repair them

The following one hundred examples are invented, not reports about actual companies or people. They are intentionally compact enough to be reused as exercises while making source meaning, reported discrepancy, evidence, correction and counterexample explicit. Each is a translation-correction investigation, not a substitute for domain-specific professional advice. The common workflow comes from the method above; the particular error and evidence are different in each case.

Finding information and navigation

Navigation reports test whether language carries the same route, destination and eligibility boundary. A fluent label is wrong if it sends the user to a different task, especially when several menu destinations have similar names. A useful way to practise is to close the proposed repair paragraph, decide independently what must survive the language change, and then compare your decision against the evidence and adversarial question.

Diagnostic 1: Request access vs get access

What the reporter sees. A menu says ‘Request access’, but users expect immediate activation after the target phrase. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source describes an application, not a guarantee. The translation implies entitlement and immediate activation. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Compare the menu label, destination heading and pending-state message. Use a target verb for applying, leaving approval conditional. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would a reader still expect to wait for approval after clicking? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 2: Return to results vs start again

What the reporter sees. A search results page has a translated ‘Back’ link that opens a new blank search. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source promises return to the existing result set. The target implies creating a new search rather than restoring context. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Save the query, pagination state and history behaviour. Align the label with the actual return action or fix the navigation. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Does the corrected label remain accurate after changing filters? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 3: View archived version

What the reporter sees. A multilingual archive link is understood as a delete function. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source offers read-only access to an older version. The target chooses a verb associated with removal rather than viewing. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Test both the label and whether the old item can be edited. Use a term for viewing historical material and preserve read-only status. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Could a careful reader believe clicking the link destroys data? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 4: Find nearby options

What the reporter sees. A service directory button translated as ‘the nearest option’ returns several possibilities. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source offers a search constrained by proximity, not a single guaranteed nearest result. The target turns a search criterion into a ranking promise. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Capture the query, results order and source design notes. Translate as discovery or nearby search without ranking certainty. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would the wording remain true if no options are found? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 5: Continue as guest

What the reporter sees. A sign-in page uses target wording suggesting an account will be created. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source permits temporary use without registration. The translation implies personal information will be stored as an account. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Reproduce with fresh browser and inspect account creation outcome. Make the guest state explicit and distinguish it from registration. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would a privacy-conscious reader understand the limited session? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Accounts, security and identity

Account text is operational language: a single verb can change a reader’s understanding of identity, session duration, credential exposure or irreversibility. Triage must include the actual authentication state. A useful way to practise is to close the proposed repair paragraph, decide independently what must survive the language change, and then compare your decision against the evidence and adversarial question.

Diagnostic 6: Sign out everywhere

What the reporter sees. A translated control resembles ‘close this window’. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source ends all active sessions for the account. The displayed target understates the cross-device effect. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Check authentication logs in a safe test account and compare session results. State explicitly that other sessions will end. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would the reader expect to remain signed in on another device? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 7: Reset password link

What the reporter sees. An email translation describes a temporary link as a new password. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source provides a link to choose a new credential. The target encourages copying a link as if it were a secret password. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Inspect template variables and the actual landing page. Differentiate resetting from receiving a password. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Could a reader disclose the link believing it is harmless? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 8: Unverified account

What the reporter sees. A profile badge translated as ‘invalid identity’ appears before email verification. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The product only knows that one verification step is incomplete. The target makes a broader judgment about the person. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Compare account data, verification requirements and badge states. Label the specific unverified attribute without moral or identity inference. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Does the badge still make sense if the user verifies by another method? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 9: Recovery code once

What the reporter sees. An instruction to save a one-time recovery code is translated as an endlessly reusable code. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source states the code can be used only once. The target loses the consumption limit. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Check the actual recovery flow and second-use error. Preserve one-time use and safe storage instructions. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Does the target warn against relying on the code a second time? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 10: Disable two-factor prompts

What the reporter sees. A settings description makes a temporary pause sound like permanent account deletion. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source changes an authentication prompt setting, not account existence. The translation attaches the action to the wrong object. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Record the toggle setting and what continues after changing it. Name the precise verification feature and its security impact. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would a cautious reader avoid the setting for the wrong reason? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Commerce, orders and payments

Commercial reports often hide mismatches between promise, displayed price, process state and later notification. The correction must protect what the source actually commits to rather than what would be reassuring. A useful way to practise is to close the proposed repair paragraph, decide independently what must survive the language change, and then compare your decision against the evidence and adversarial question.

Diagnostic 11: Authorization vs charge

What the reporter sees. At checkout a translated notice says the card has been charged while the processor only placed a hold. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source describes authorization, with settlement still pending. The target confuses a temporary hold with completed collection. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Compare payment gateway state, receipt and bank-facing copy. Use the approved concept for authorization and avoid promising release dates. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Does the notice remain true if the hold never settles? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 12: Refund issued vs requested

What the reporter sees. A customer receives an email that translates ‘refund requested’ as ‘refund completed’. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source records the start of a process. The target promises money already returned. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Reproduce with a pending refund in the sandbox. Differentiate requested, approved, initiated and completed states. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Could a customer reasonably stop following up after this notice? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 13: Discount up to

What the reporter sees. A promotion banner loses ‘up to’ before a percentage. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source gives a maximum possible discount. The target makes the maximum sound universal. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Review the offer rules and several eligible item prices. Restore the upper-bound language and placement. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Does a low-discount item make the translated banner misleading? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 14: Delivery estimate

What the reporter sees. A shipping message turns ‘estimated arrival’ into ‘guaranteed arrival’. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source states a forecast with uncertainty. The target changes a forecast to a contractual certainty. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Compare logistics state and approved service commitments. Use the target language’s normal uncertainty marker. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would a delay contradict a promise the company never made? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 15: Return window starts

What the reporter sees. A returns FAQ is translated with a window starting at purchase rather than receipt. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source ties timing to delivery confirmation. The target shortens or extends actual eligibility. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Model orders where purchase and delivery dates differ. Name the triggering event accurately. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would a late-delivered item be unfairly excluded under the wording? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Consent, permissions and data use

Permission and privacy language must distinguish consent, required processing, optional settings and mere notification. Avoid making legal claims from an unchecked source; escalate unresolved scope to the authorised policy owner. A useful way to practise is to close the proposed repair paragraph, decide independently what must survive the language change, and then compare your decision against the evidence and adversarial question.

Diagnostic 16: Accept essential cookies

What the reporter sees. A banner translates a technical necessity as blanket consent to all tracking. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source distinguishes essential operation from optional measurement. The target combines two different purposes. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Inspect the consent categories, default values and actual switches. Translate each category by its intended processing purpose. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Can the reader refuse optional tracking without misunderstanding? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 17: Withdraw permission

What the reporter sees. A help text implies withdrawal deletes historical records automatically. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source describes ending a particular future permission. The target adds a deletion consequence not supported by the policy. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Consult the source owner and inspect account state transitions. Preserve withdrawal as a specific action without invented effects. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would the reader believe older copies must instantly disappear? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 18: Share with collaborators

What the reporter sees. A document prompt translated as ‘publish publicly’ appears for restricted sharing. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source grants access to named collaborators only. The target confuses invitation and public disclosure. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Test with an invited user and an anonymous visitor. Describe the actual audience and permission scope. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Can a person outside the invited list open the document? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 19: Preference saved locally

What the reporter sees. A privacy note translates ‘stored on this device’ as ‘sent to our server’. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source limits storage location. The target reverses a material privacy fact. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Verify engineering documentation and data-flow behaviour. Correct location language after confirming the underlying system. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would the same text survive a browser storage reset? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 20: Optional field vs mandatory

What the reporter sees. A registration form’s translated label marks optional demographic data as required. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source allows the user to skip the field. The target suggests consent is compulsory. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Submit with the field empty and examine validation output. Restore optionality in label, hint and error states. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Does the person successfully finish without disclosing the data? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Health and wellbeing communications

Health-related content requires special care: the translation process may be explained, but clinical interpretation and individual advice require qualified professionals. Preserve measured facts, instruction hierarchy and urgency without generating new medical advice. A useful way to practise is to close the proposed repair paragraph, decide independently what must survive the language change, and then compare your decision against the evidence and adversarial question.

Diagnostic 21: Take with food

What the reporter sees. A leaflet translation suggests taking a product before any meal, even when the source says with food. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source specifies administration alongside food. The target changes timing. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Check the authorised product leaflet and approved medical wording. Escalate to a qualified reviewer and restore the source timing. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would a reader choose a different time using the target alone? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 22: Possible side effect

What the reporter sees. A patient information page turns ‘may cause drowsiness’ into ‘will cause drowsiness’. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source communicates possibility rather than certainty. The target overstates incidence and certainty. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Compare the authorised statement and modal wording. Keep the uncertainty marker and avoid invented probabilities. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Can a reader distinguish risk from inevitability? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 23: Do not exceed

What the reporter sees. A translated label loses negation before a maximum quantity. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source sets an upper limit. The target appears to recommend crossing that limit. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Record package version and all distribution channels immediately. Contain the unsafe display and refer correction for qualified approval. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would the corrected target prevent the prohibited action? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 24: Appointment confirmed

What the reporter sees. A reminder translates ‘appointment request received’ as an accepted appointment. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source acknowledges receipt only. The target implies scheduling confirmation. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Test pending and confirmed appointment states. Use separate translations for requested and confirmed stages. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Will the reader still verify the final appointment time? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 25: Screening vs diagnosis

What the reporter sees. An educational article translates screening result as a confirmed diagnosis. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source describes an indicator needing further assessment. The target creates unwarranted clinical certainty. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Consult the authoritative screening definitions and context. Restore the screening status and appropriate limitation. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would the target remain accurate for a false-positive result? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Warnings and critical instructions

For critical instructions, an error report can trigger immediate containment before the source and translation investigation is complete. The aim is preserving a safe action, not generating improvised emergency directions. A useful way to practise is to close the proposed repair paragraph, decide independently what must survive the language change, and then compare your decision against the evidence and adversarial question.

Diagnostic 26: Do not enter

What the reporter sees. A restricted doorway label loses negation in one target. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source forbids entry during a hazardous state. The target reverses a prohibition. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Capture the exact sign location, active hazard and approved wording. Stop distribution if warranted and use reviewed safety language. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would an uninformed reader still avoid entering? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 27: Wait for all-clear

What the reporter sees. An alert translated as ‘return immediately’ appears before an all-clear. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source requires waiting for an explicit safe-to-return notice. The target inverts the chronological condition. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Compare current alert status, issued revisions and source authority. Correct the condition under responsible incident command. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would a person return prematurely after reading it? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 28: Ventilation instruction

What the reporter sees. A translated equipment warning mistakes ‘do not block’ for ‘close’. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source requires keeping vents unobstructed. The target changes the required physical state. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Check equipment documentation and label placement. Use approved technical wording and recheck visual symbols. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Could the person deliberately obstruct the airflow? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 29: Authorized personnel only

What the reporter sees. A translated access sign says visitors must pass through a restricted area. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source limits access to authorised staff. The target turns restriction into route guidance. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Reproduce on the exact sign and map sequence. Restore the access restriction and correct any conflicting arrows. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would the corrected sign keep visitors outside the area? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 30: Stop when alarm sounds

What the reporter sees. A safety instruction translates ‘stop operation’ as ‘stop alarm’. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source concerns machinery operation, not silencing the warning. The target points action at the wrong object. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Consult the technical operator and approved safety procedure. Name the required machine action precisely. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Could the reader silence an alarm while leaving the machine running? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Banking and statements

Financial interfaces create semantic traps around balance, availability, settlement, estimated rates and disputed charges. A correction needs the finance-state owner, not just an elegant target phrase. A useful way to practise is to close the proposed repair paragraph, decide independently what must survive the language change, and then compare your decision against the evidence and adversarial question.

Diagnostic 31: Available vs current balance

What the reporter sees. A translated account field labels available funds as total balance. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source distinguishes spendable funds from posted balance. The target makes different accounting measures interchangeable. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Compare transaction holds and example account states. Name the field according to its actual calculation. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Does a pending debit make the difference visible? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 32: Pending transfer

What the reporter sees. A notification tells the reader a transfer arrived when it is still processing. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source reports pending transfer status. The target makes processing sound completed. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Review payment lifecycle stages and source event keys. Translate pending, accepted and credited as distinct states. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would the user rely on funds not yet credited? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 33: Indicative rate

What the reporter sees. A currency quote is translated as the final applied rate. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source offers an estimate subject to execution time. The target erases the quotation caveat. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Compare transaction timestamps and final receipts. Restore the estimate qualifier and relevant condition. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Can the execution rate differ without contradicting the label? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 34: Minimum payment

What the reporter sees. A credit statement translates minimum due as the full outstanding balance. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source identifies the minimum required payment. The target confuses one threshold with total liability. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Check statement field definitions and policy approval. Use the established finance term without changing the amount. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would the reader understand the remaining balance still exists? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 35: Disputed transaction

What the reporter sees. A card dispute acknowledgement is translated as a confirmed fraud decision. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source only acknowledges an allegation under review. The target makes an unresolved investigation final. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Check dispute stage and event sequence. Keep allegation, review and outcome separate. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would the statement still be true if the claim is rejected? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Workplace and organisation

Workplace translation must preserve authority, responsibility, task status and boundaries between instruction and suggestion. Resolve management policy at its owner rather than making the target language more forceful. A useful way to practise is to close the proposed repair paragraph, decide independently what must survive the language change, and then compare your decision against the evidence and adversarial question.

Diagnostic 36: Approval required

What the reporter sees. A workflow notice translated as ‘approval recommended’ appears on a mandatory control. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source makes approval a precondition. The target weakens the obligation. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Test whether the process blocks without approval. Preserve required status and authorised approver role. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Can the user proceed lawfully without the stated approval? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 37: Draft policy

What the reporter sees. A document header translated as ‘official policy’ marks a draft. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source says the content is not final. The target implies immediate institutional force. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Locate version history and adoption record. Keep draft status visible in all target variants. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would staff treat an unapproved change as binding? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 38: Manager review

What the reporter sees. A message says ‘review pending’ but target suggests the manager has already approved. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source marks an outstanding action. The target changes process stage and expectation. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Inspect workflow timestamps and role labels. Distinguish waiting, reviewing and approved states. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would the requester bypass a necessary review? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 39: May work remotely

What the reporter sees. A guidance note is translated as an entitlement rather than discretionary permission. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source allows remote work subject to conditions. The target removes the stated conditional authority. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Check the relevant employee policy and approval rule. Retain modal force and conditional language. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would an employee believe no request is necessary? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 40: Incident near miss

What the reporter sees. A safety form calls a near miss a completed injury. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source records a hazardous event without actual injury. The target changes a reporting category. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Compare event taxonomy and form submission outcomes. Use a recognised category with an explanatory hint if needed. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Could the change distort statistical reporting? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Learning and assessment

Educational content can fail when target wording changes what learners are asked to do, the level of certainty, time allowed or the rubric criterion. The correction must keep the task equivalent without copying the English-learning owner’s curriculum. A useful way to practise is to close the proposed repair paragraph, decide independently what must survive the language change, and then compare your decision against the evidence and adversarial question.

Diagnostic 41: Select two answers

What the reporter sees. An assessment prompt omits the number two. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source requires exactly two selections. The target allows an unspecified number. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Try the test with one, two and three choices in the same state. Restore the quantity and verify input validation. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Does the instruction agree with the scoring rule? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 42: Explain vs list

What the reporter sees. A translated question says list examples when the source asks to explain causes. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source assesses causal reasoning. The target substitutes recall for explanation. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Review question purpose, rubric and sample responses. Use a target task verb that elicits explanation. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would a correct list-only answer satisfy the actual rubric? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 43: Open-book rule

What the reporter sees. A test direction says no notes where the source permits approved reference materials. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source allows specified materials. The target removes permitted support. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Read the authorised assessment policy and proctor instruction. Preserve the exact scope of permitted resources. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would the candidate prepare differently because of the translation? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 44: Deadline timezone

What the reporter sees. An assignment page translates the due date without its time-zone context. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source defines a timestamp rather than a universal local wall clock time. The target invites inconsistent interpretations across regions. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Compare the underlying timestamp and displayed locale format. Specify zone or equivalent absolute time unambiguously. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Could two readers in different zones infer different deadlines? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 45: Feedback not grade

What the reporter sees. A learner dashboard translates formative commentary as a final grade. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source labels feedback designed for improvement. The target signals a summative decision. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Check whether the assessment is recorded in the gradebook. Separate developmental feedback from graded status. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would the learner believe the result cannot change? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

News, public explanation and civic content

Reports about news translation must preserve who said what, whether a claim was verified, and the status of public processes. Editing language must never silently convert allegations into findings. A useful way to practise is to close the proposed repair paragraph, decide independently what must survive the language change, and then compare your decision against the evidence and adversarial question.

Diagnostic 46: Alleged vs proved

What the reporter sees. A report translates ‘alleged misconduct’ as established guilt. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source attributes an unproven claim. The target removes the evidence-status distinction. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Examine the original attribution and publication date. Restore allegation and identification of the speaker. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Could the target cause a reader to treat an accusation as a finding? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 47: Preliminary results

What the reporter sees. An election-related information label turns preliminary counts into final certification. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source states a provisional status. The target implies official completion. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Consult the relevant issuing authority and revision history. Maintain the distinction between projected, counted and certified. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would later official changes look like contradiction? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 48: Reportedly vs definitely

What the reporter sees. A news brief loses a reporter-attribution marker. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source presents information as reported by a specified party. The target makes the publisher assert it directly. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Compare quotations, sourcing and original evidence. Reinsert attribution in natural target form. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Can readers identify who is responsible for the claim? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 49: Historical date uncertainty

What the reporter sees. An archive caption translates ‘circa’ as an exact date. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source describes an approximate period. The target adds precision not supported by records. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Check catalogue metadata and provenance notes. Use an accepted approximation marker. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would the translation falsely certify an exact day? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 50: Public consultation open

What the reporter sees. A notice translated as an enacted rule appears while comments are still invited. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source describes a proposal under consultation. The target says the policy already applies. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Check proposal status, deadline and authority record. State the current procedural stage accurately. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would citizens believe feedback is no longer possible? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Media, subtitles and audio

Media translation failures may occur in a specific timestamp, audio track, speaker identity or reading time. The correction record must name the actual cue or track and coordinate with its dedicated audiovisual owner. A useful way to practise is to close the proposed repair paragraph, decide independently what must survive the language change, and then compare your decision against the evidence and adversarial question.

Diagnostic 51: Subtitle speaker swapped

What the reporter sees. A documentary subtitle assigns a quotation to the narrator rather than a guest. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source voice belongs to the guest. The target suggests a different person made the claim. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Compare audio, cue timings and speaker identifiers. Correct attribution and test the released rendition. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would viewers still know who made the statement? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 52: Caption drops negation

What the reporter sees. An instructional video caption omits ‘not’ because the line was shortened. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source forbids a particular action. The target appears to recommend it. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Locate the exact cue identifier and time offset. Restore meaning while resegmenting within readable constraints. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Is the prohibition legible for the whole relevant action? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 53: Dubbing wrong tense

What the reporter sees. A dubbed explanation describes a completed event as ongoing. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The original voice presents a finished historical action. The target changes chronology. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Compare script, recorded dub and final mix. Record a corrected line with compatible timing. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Could a viewer interpret the event as still occurring? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 54: Audio description role

What the reporter sees. A translated audio description names an object that the scene does not confirm. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source description expresses uncertainty. The target turns inference into fact. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Inspect video frames and the approved description script. Restore the uncertainty rather than adding an invented identity. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Could a blind viewer distinguish seen fact from guess? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 55: Content warning scope

What the reporter sees. A translated content notice expands a mild thematic warning into a graphic-content promise. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source warns about a topic, not particular visual intensity. The target changes audience expectation. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Review scene-level evidence and approved content labels. Use accurate scope and avoid unsupported detail. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would a viewer make a different informed viewing decision? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Software, developer interfaces and APIs

Technical strings can be perfectly grammatical while lying about state transitions, parameter values or system guarantees. Developers should give translators resource identifiers and functional context, then test the released machine-visible boundary. A useful way to practise is to close the proposed repair paragraph, decide independently what must survive the language change, and then compare your decision against the evidence and adversarial question.

Diagnostic 56: Request queued

What the reporter sees. An API console translates ‘queued’ as ‘executed’. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source indicates accepted scheduling, not completed work. The target makes queued work look final. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Compare job IDs and asynchronous lifecycle events. Use separate target labels for scheduled and completed. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Will the state be true before the worker runs? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 57: Retry safely

What the reporter sees. An error page instructs the user to repeat a non-idempotent operation. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source warns not to submit again until status is known. The translated help text encourages duplicate execution. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Simulate network interruption and check duplicate record risk. Preserve the non-retry condition, with technical owner review. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Could following the translation trigger a second charge or order? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 58: Delete cache vs records

What the reporter sees. A maintenance dialog uses the same translated noun for temporary cache and saved records. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source distinguishes disposable performance data from durable user content. The target hides the irreversible difference. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Test cleanup effects and field definitions. Name the exact object and permanence. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would the reader consent to data loss they did not anticipate? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 59: Rate limit

What the reporter sees. A developer message translates ‘requests per minute’ as total requests allowed. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source defines a time-window threshold. The target removes the denominator. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Check documented limit scope and counter reset behaviour. Restore the rate and its unit. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would a developer design a wrong capacity plan? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 60: Stale webhook

What the reporter sees. A warning translated as ‘invalid signature’ appears for a delivery older than an allowed age. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source says the event was rejected for timing. The target falsely blames cryptographic authentication. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Use controlled expired and invalid-signature test events. Assign separate messages to distinct failure modes. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Can operators diagnose the right cause from the target? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Assistive interaction and accessible reports

Accessibility-linked reports must be reproducible through the actual assistive pathway. The issue may live in hidden text, status announcement or language tagging rather than the visible label. A useful way to practise is to close the proposed repair paragraph, decide independently what must survive the language change, and then compare your decision against the evidence and adversarial question.

Diagnostic 61: Screen reader says wrong language

What the reporter sees. A visible target sentence is correct, but a screen reader pronounces it as another language. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The string’s language must be available to assistive technology. The page’s language annotation is wrong. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Capture DOM language values and the spoken result with a named setup. Correct the language metadata at the owning component. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Does the fix work without requiring the user to change device language? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 62: Icon-only report control

What the reporter sees. A translation-report icon lacks an accessible name in one locale. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The control should communicate that it opens a language-feedback form. The visible icon alone is ambiguous to assistive users. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Inspect focus order and accessibility tree. Provide a meaningful target-language accessible name. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Can a keyboard and screen-reader user find and activate it? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 63: Error summary missing field

What the reporter sees. A feedback form translated its required-field error as ‘something is wrong’. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source identifies the particular missing field. The target removes actionable detail. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Submit the form with several different empty fields. Name the invalid field in target language and associate the message. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Can the reporter correct the error without guessing? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 64: Success announcement too early

What the reporter sees. An accessible live region translates ‘upload started’ as ‘report submitted’. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source marks an in-progress operation. The target provides false completion feedback. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Test network delay and upload failure states. Use stage-specific announcements. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Will a user know when a report actually arrived? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 65: Language switch preserves form

What the reporter sees. Changing feedback form language silently clears the typed problem description. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The interface promises language switching, not loss of draft input. The process invalidates the reader’s effort and weakens reporting access. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Reproduce with long unsent text and different languages. Fix state persistence and translate the status notice accurately. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Can the user return without losing their words? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Community, platforms and moderation

Community interactions involve reporting, accusation, visibility and appeal. The meaning of a translation may affect someone’s reputation or ability to challenge a decision. A useful way to practise is to close the proposed repair paragraph, decide independently what must survive the language change, and then compare your decision against the evidence and adversarial question.

Diagnostic 66: Flag for review

What the reporter sees. A user report button translated as ‘ban user’ appears before moderation. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source submits a report for evaluation. The target implies immediate punishment. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Test the real moderation workflow and submitter notification. Use a verb for raising a concern, not applying a sanction. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would the reporter think they control the outcome? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 67: Post hidden pending review

What the reporter sees. A message says a post was deleted when it is temporarily hidden. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source describes a reversible visibility state. The target implies permanent removal. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Check moderation state and appeal availability. Separate hidden, removed and deleted. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Could a creator wrongly believe restoration is impossible? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 68: Appeal accepted

What the reporter sees. An appeal acknowledgement is translated as a successful appeal decision. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source confirms intake only. The target signals a favourable final outcome. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Test appeal stages and notifications. Translate submitted, accepted for review and granted separately. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would the person expect a restriction to lift immediately? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 69: Private group visibility

What the reporter sees. A post label translated as public appears in a closed group. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source restricts the audience to members. The target changes the privacy expectation. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Verify actual read permissions across roles. Name the precise sharing scope after technical confirmation. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Could the user disclose something they otherwise would not? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 70: Quote attribution

What the reporter sees. Automated translation of an embedded quote makes it appear to be the poster’s original claim. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source marks another speaker’s words as quotation. The target loses attribution boundaries. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Inspect original markup, speaker and rendered thread. Preserve quotation and speaker identity. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Could the reader assign responsibility to the wrong person? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Travel, transport and mobility

Timetables, routes and travel notices depend on states and time. Translation corrections should preserve the difference between an estimate, an instruction and an official service update. A useful way to practise is to close the proposed repair paragraph, decide independently what must survive the language change, and then compare your decision against the evidence and adversarial question.

Diagnostic 71: Platform change

What the reporter sees. A station notice translates ‘platform subject to change’ as a confirmed platform. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source indicates provisional allocation. The target removes conditional status. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Compare live timetable, announcement version and display timestamp. Restore provisional wording and current channel links. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would someone assume the platform cannot change? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 72: Boarding closing

What the reporter sees. A boarding reminder translates ‘boarding closes’ as ‘departure time’. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source identifies an earlier deadline. The target shifts the relevant action window. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Model itinerary times with a gap between closing and departure. Use the exact process event. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Could a traveller arrive too late despite following the translation? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 73: Service suspended

What the reporter sees. A timetable banner translated as ‘service delayed’ appears during a suspension. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source says operation has stopped. The target invites indefinite waiting. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Check service state and official alert revision. Distinguish suspension from delay. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would the reader plan around a train that will not operate? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 74: Transfer required

What the reporter sees. A route planner translation says direct journey where a transfer is required. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source specifies changing vehicles. The target conceals an itinerary step. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Compare leg sequence and route display. Make the transfer visible and consistent across views. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Can an unfamiliar reader complete the trip correctly? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 75: Estimated gate open

What the reporter sees. A travel app translated ‘estimated gate opening’ as current boarding permission. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source forecasts a future opening time. The target tells the reader to act too soon. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Compare operational events and published timing. Keep forecast and present state separate. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would the message remain true if staff open the gate late? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Science, measurement and evidence

Scientific correction work must preserve measured quantity, uncertainty, method and category. A plausible-looking target number can be wrong even when every digit is copied exactly. A useful way to practise is to close the proposed repair paragraph, decide independently what must survive the language change, and then compare your decision against the evidence and adversarial question.

Diagnostic 76: Confidence interval

What the reporter sees. A translated chart caption describes a confidence interval as the observed data range. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source describes an inferential interval. The target redefines the statistical object. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Compare study method, axes and source terminology. Choose the accepted target term with context. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would the caption still be true if observations lie outside the interval? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 77: Median vs average

What the reporter sees. A report translates median as arithmetic mean. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source identifies the middle ranked value. The target changes the statistical measure. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Compare the data distribution and calculation definitions. Restore the correct measure name. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would a skewed dataset expose the difference? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 78: Below detection limit

What the reporter sees. A laboratory summary translates ‘below detection’ as ‘zero’. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source says the method did not detect a quantity above threshold. The target claims none exists. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Review assay detection limit and reporting convention. Keep non-detection distinct from zero. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would a more sensitive instrument potentially find a value? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 79: Forecast vs observed

What the reporter sees. A climate chart target legend calls model projection a measurement. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source labels modelled future values. The target changes evidence type. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Check dataset variables, scenarios and time ranges. Preserve projection status and conditional assumptions. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Could a reader cite the projection as a measured historical fact? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 80: Units per volume

What the reporter sees. A target label preserves ‘5’ but swaps milligrams per litre for milligrams per millilitre. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source’s concentration includes a specific denominator. The target changes magnitude by orders of scale. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Compare approved units, data schema and chart annotations. Correct the complete quantity unit and test conversions. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would a numeric-only comparison detect the error? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Food, packaging and product information

Consumer labels combine names, contents, restrictions, process conditions and dates. The correction must use controlled product facts and should never infer a safety claim from machine translation alone. A useful way to practise is to close the proposed repair paragraph, decide independently what must survive the language change, and then compare your decision against the evidence and adversarial question.

Diagnostic 81: May contain allergen

What the reporter sees. A translated label states ‘contains’ where the source uses precautionary ‘may contain’. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source expresses potential cross-contact, not a confirmed ingredient. The target changes claim strength. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Compare approved label and allergen statement category. Restore the precise qualified warning via authorised review. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Could a buyer understand confirmed ingredient presence rather than possible risk? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 82: Use by vs best before

What the reporter sees. A package label translates a quality date as a safety deadline or vice versa. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source distinguishes different date functions. The target changes the type of consumer guidance. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Inspect product classification and approved original artwork. Use the correct date category and local labelling rules where applicable. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would the buyer treat the date as an entirely different kind of limit? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 83: Keep refrigerated

What the reporter sees. A storage instruction translates ‘after opening’ as ‘before opening’. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source ties refrigeration to a change in package state. The target reverses the condition. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Compare product specification and storage states. Restore the conditional sequence under product owner approval. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would someone store the opened item incorrectly? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 84: Net weight vs drained

What the reporter sees. A label translates net drained weight as total net weight. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source measures product after liquid removal. The target changes the measurement basis. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Inspect packaging data and measurement methods. Retain the distinct measured quantity. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would customers compare unlike amounts between products? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 85: Product variant naming

What the reporter sees. A product page translates a variant name into a different formulation. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source identifies a particular SKU and ingredients. The target swaps product identity. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Compare identifiers, packaging, ingredients and catalogue mapping. Correct the catalogue-to-language mapping before style polish. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Could a person order a different item from the one shown? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Support, claims and customer disputes

Support communication records what has happened, what is proposed and which path remains open. An incorrect translation can inadvertently close a complaint or create an obligation not supported by the case. A useful way to practise is to close the proposed repair paragraph, decide independently what must survive the language change, and then compare your decision against the evidence and adversarial question.

Diagnostic 86: Ticket received

What the reporter sees. A support portal translated ‘request received’ as ‘problem solved’. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source acknowledges intake. The target implies remedy already completed. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Reproduce initial ticket and resolved ticket states. Use distinct accepted and resolved target terms. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would a customer stop seeking help prematurely? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 87: Replacement offered

What the reporter sees. A reply says the customer was refunded though a replacement is only being offered. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source proposes one remedy, not completed payment. The target changes both remedy type and completion. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Check case notes and authorised offers. Restore the conditional replacement offer. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Could the customer rely on money never returned? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 88: Case closed administratively

What the reporter sees. A message states ‘claim rejected’ though the ticket was closed because of duplication. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source describes administrative consolidation. The target gives a false merits decision. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Compare linked cases and closure reason codes. Explain that the report was merged rather than denied. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would the person incorrectly think appeal rights are exhausted? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 89: Further documents requested

What the reporter sees. A translated support message says ‘approved’ although additional evidence is required. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source makes next action the customer’s responsibility. The target removes the outstanding requirement. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Review workflow state and checklist. Translate the evidence request precisely and use action-oriented wording. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Will the reader know what must be provided? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 90: Goodwill offer

What the reporter sees. A voluntary credit is translated as a statutory entitlement. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source describes discretionary support. The target asserts a legal right not verified by the issuer. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Consult policy owner and offer terms. Preserve offer status without unnecessary legal claims. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Could later customers assume the credit is automatically owed? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Machines, devices and hardware

Hardware reporting must distinguish commands, states, physical units and irreversible actions. Any apparent safety risk should be referred to the qualified responsible technical owner. A useful way to practise is to close the proposed repair paragraph, decide independently what must survive the language change, and then compare your decision against the evidence and adversarial question.

Diagnostic 91: Standby vs off

What the reporter sees. A device menu translates standby as powered off. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source keeps some systems energised. The target implies complete disconnection. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Test the specified device state safely and inspect the manual. Restore distinct operational modes. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Could a user assume no electrical power remains? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 92: Factory reset

What the reporter sees. A translated warning says settings will restart rather than data will be erased. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source describes an operation with data-loss implications. The target hides irreversibility. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Test on non-production hardware and verify reset scope. Describe settings reset and any data deletion accurately. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would a user back up important information before acting? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 93: Calibration required

What the reporter sees. A dashboard translates calibration needed as calibration completed. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source indicates work remains outstanding. The target gives false confidence in measurement quality. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Inspect state flags and calibration records. Use a pending-action message with clear owner. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Could an operator trust uncalibrated readings? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 94: Rotate clockwise

What the reporter sees. A translated adjustment instruction reverses rotation direction. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source describes a specific physical motion. The target changes the action. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Compare official diagram, perspective and part numbering. Use a verified target-language direction with visual alignment. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Will the observed hardware movement match the diagram? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 95: Fault cleared vs muted

What the reporter sees. An alert reports muted alarms as repaired faults. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source describes sound suppression, not fault resolution. The target confuses notification state with system health. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Test mute and clear states separately. Name the exact state and retain unresolved issue indication. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Could staff leave a real fault unattended? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Archives, scholarship and cultural records

Archival translations should preserve provenance, attribution, uncertain identity and historical framing. A correction can be necessary precisely because the target text sounds more certain than the evidence. A useful way to practise is to close the proposed repair paragraph, decide independently what must survive the language change, and then compare your decision against the evidence and adversarial question.

Diagnostic 96: Attributed to

What the reporter sees. A catalogue translates ‘attributed to’ as ‘made by’. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source explicitly marks uncertain authorship. The target asserts a confirmed creator. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Review acquisition records and catalogue status. Use an attribution marker that retains uncertainty. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would new evidence contradict the claim of certain authorship? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 97: Date range

What the reporter sees. A manuscript record translates a broad century range into a single year. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source contains only an approximate period. The target invents specificity. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Compare original catalogue fields and dating evidence. Restore the period or accepted approximation. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would a historian cite an unsupported precise date? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 98: Former collection name

What the reporter sees. An archive target silently replaces a historical institution name with its modern successor. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source records the institution at a particular time. The target changes provenance context. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Inspect accession history and naming timeline. Keep historical name with an explanatory modern link if justified. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Could the reader misidentify who owned the object then? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 99: Partial transcription

What the reporter sees. A manuscript description translates ‘partially legible’ as complete transcription. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source acknowledges missing or unclear text. The target claims comprehensive recovery. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Check image quality and transcription annotations. Restore partial-status language and uncertainty. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Would scholars assume missing passages were read? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Diagnostic 100: Unverified oral testimony

What the reporter sees. A record translated as an official finding comes from an interviewee’s recollection. The reporter’s statement is an observation to investigate, not yet a verdict about who made a mistake. First preserve the exact displayed target and the context in which a person was likely to rely on it. Locate the source text, the target-language version, the release channel and the functional state involved.

Meaning contract and defect. The source is attributed personal testimony. The target erases source type. The decisive question is not whether every word has a dictionary equivalent. It is whether an ordinary target-language reader would draw the same consequential inference and carry out the same relevant action as the source permits. Keep the reported interpretation distinct from what the product or document actually does.

Reproduce, decide and correct. Review interview permission and citation details. Identify the speaker and status accurately. Before accepting a proposed new phrase, ask the authorised owner to confirm the relevant fact, condition or action. Then check the candidate target independently for naturalness, and compare it against the original for omissions, additions, modality, scope, reference and timing. The correction record should identify its approved source version.

Adversarial revalidation. Could readers mistake memory for independently verified fact? Run the counterexample in the released target interface or document, not just the translator’s preview. If the wording is still misleading in that state, the issue is not verified. Note which related locale variants, notifications, offline copies or assistive surfaces were also tested; record any channel that remains unknown rather than suggesting a universal fix.

Twenty applied workshops: from recognition to independent correction

The workshops are designed to be used with fictional or appropriately anonymised material. Complete the diagnosis before reading the model prompts, and compare the result with a specific acceptance condition. Do not use real customer secrets, private patient notes or confidential source files in a public exercise.

Practice A. Rebuild an incomplete report

You receive a complaint consisting only of ‘bad translation’ and a cropped image of a button. Instead of requesting the reporter’s identity, draft a compassionate two-question follow-up: where did you see this, and what did you expect it to mean or do? Next identify a way to recover the content key internally without burdening the reporter. Show how the same evidence would distinguish a wrong verb from a destination mapping problem. The practical skill is extracting enough evidence to reproduce the claim while collecting no unnecessary private data.

A strong submitted answer should include a one-sentence meaning contract, the observed released behaviour, a distinct hypothesis about the delivery path, a named authorised decision maker and one counterexample that would falsify the proposed correction. A weak answer merely offers a more elegant sentence. When the information is incomplete, the correct professional response is to describe the missing evidence and a safe route for obtaining it, not to guess.

Practice B. Rank a stylistic complaint against a safety reversal

A home page headline uses an awkward but comprehensible metaphor. A device warning in a low-traffic language has lost a prohibition. Explain why complaint popularity, visual prominence and severity may point in different directions. Construct a triage note that names consequence, exposure, confidence and next action. The answer should provide temporary containment for the dangerous variant while preserving a normal editorial queue for the headline. Avoid treating a low number of users as evidence of low individual risk.

A strong submitted answer should include a one-sentence meaning contract, the observed released behaviour, a distinct hypothesis about the delivery path, a named authorised decision maker and one counterexample that would falsify the proposed correction. A weak answer merely offers a more elegant sentence. When the information is incomplete, the correct professional response is to describe the missing evidence and a safe route for obtaining it, not to guess.

Practice C. Detect a hidden false guarantee

A translated notification says ‘completed’, but the actual system state is ‘queued’. Trace the message across source copy, translation key, event type, notification service and rendered history. Write an acceptance test that includes at least one pending state and one completed state. The test should fail if either state shows the same misleading label. Explain why a fluent sentence evaluated without application context could receive a good language score yet remain a defective user instruction.

A strong submitted answer should include a one-sentence meaning contract, the observed released behaviour, a distinct hypothesis about the delivery path, a named authorised decision maker and one counterexample that would falsify the proposed correction. A weak answer merely offers a more elegant sentence. When the information is incomplete, the correct professional response is to describe the missing evidence and a safe route for obtaining it, not to guess.

Practice D. Compare source and target without blindly preferring literal language

A source message says ‘may receive a reply within two days’. The first translation sounds natural but says the reply will arrive; the second uses awkward syntax yet retains possibility. Draft a third option that restores uncertainty in natural target-language form. Explain exactly what the source does not promise. Add a test using a scenario in which no reply arrives. If the translated notice now constitutes a broken guarantee, it remains wrong.

A strong submitted answer should include a one-sentence meaning contract, the observed released behaviour, a distinct hypothesis about the delivery path, a named authorised decision maker and one counterexample that would falsify the proposed correction. A weak answer merely offers a more elegant sentence. When the information is incomplete, the correct professional response is to describe the missing evidence and a safe route for obtaining it, not to guess.

Practice E. Decide when the source needs amendment

A multilingual FAQ describes a cancellation period using the phrase ‘by Friday’ but the business rule is coded as Friday at 00:00. Do not select a target-language preposition by intuition. Identify the missing source-policy precision and obtain an authorised rule. Propose a revised source containing a clear time zone and cut-off point, then translate it. State why the translator should not resolve policy uncertainty by inventing a deadline.

A strong submitted answer should include a one-sentence meaning contract, the observed released behaviour, a distinct hypothesis about the delivery path, a named authorised decision maker and one counterexample that would falsify the proposed correction. A weak answer merely offers a more elegant sentence. When the information is incomplete, the correct professional response is to describe the missing evidence and a safe route for obtaining it, not to guess.

Practice F. Inspect a fallback path

A French-language app has a corrected target string in the localisation editor. A regional variant inherits an older translation from a fallback bundle. Map language preference, exact regional variant, general-language fallback and default-language behaviour. Check on a clean device and a previously used device. Record which release versions carry the correction. Explain why a successful editor screenshot is insufficient proof that the affected users see the new text.

A strong submitted answer should include a one-sentence meaning contract, the observed released behaviour, a distinct hypothesis about the delivery path, a named authorised decision maker and one counterexample that would falsify the proposed correction. A weak answer merely offers a more elegant sentence. When the information is incomplete, the correct professional response is to describe the missing evidence and a safe route for obtaining it, not to guess.

Practice G. Write a corrective customer reply

A user reports that a refund email was translated as ‘refund paid’ when the finance system recorded only ‘request submitted’. Write a short response acknowledging the verified wording defect while avoiding a promise about payment timing. Include what was changed and how to locate the updated status. Keep private financial information out of the public explanation. Show how you would handle a user who received the misleading version before the fix.

A strong submitted answer should include a one-sentence meaning contract, the observed released behaviour, a distinct hypothesis about the delivery path, a named authorised decision maker and one counterexample that would falsify the proposed correction. A weak answer merely offers a more elegant sentence. When the information is incomplete, the correct professional response is to describe the missing evidence and a safe route for obtaining it, not to guess.

Practice H. Audit one AI-generated patch

An AI system proposes replacing a reported term with a phrase that is more fluent. Demand a source link, original context, locale, human review, domain meaning and test. Check whether the proposal broadens an exception or strengthens a claim. Record prompt or model configuration only where it is necessary for reproducibility. Explain why the AI’s confident explanation cannot be treated as source evidence, even if its translation sounds professional.

A strong submitted answer should include a one-sentence meaning contract, the observed released behaviour, a distinct hypothesis about the delivery path, a named authorised decision maker and one counterexample that would falsify the proposed correction. A weak answer merely offers a more elegant sentence. When the information is incomplete, the correct professional response is to describe the missing evidence and a safe route for obtaining it, not to guess.

Practice I. Verify a report from an assistive user

A screen-reader user reports wrong pronunciation for a sentence that looks correct on screen. Document the language-tag context, accessible name or text node, screen-reader configuration and actual spoken result. Change only the relevant component in a test environment. Ask the reporter for confirmation if feasible, but do not make their participation the sole proof of accessibility. Ensure the fixed version works without exposing the user’s device details.

A strong submitted answer should include a one-sentence meaning contract, the observed released behaviour, a distinct hypothesis about the delivery path, a named authorised decision maker and one counterexample that would falsify the proposed correction. A weak answer merely offers a more elegant sentence. When the information is incomplete, the correct professional response is to describe the missing evidence and a safe route for obtaining it, not to guess.

Practice J. Distinguish translation errors from authorised variants

Two reviewers disagree over formal and informal second-person address in a target language. Gather audience, relationship, brand voice and source purpose. Decide whether either variant changes respect, obligation or implied social distance materially. If both are valid, record a preferred style rule rather than escalating a false critical incident. Explain why legitimate community differences require evidence and consultation rather than a single universal claim.

A strong submitted answer should include a one-sentence meaning contract, the observed released behaviour, a distinct hypothesis about the delivery path, a named authorised decision maker and one counterexample that would falsify the proposed correction. A weak answer merely offers a more elegant sentence. When the information is incomplete, the correct professional response is to describe the missing evidence and a safe route for obtaining it, not to guess.

Practice K. Revalidate an archived or printed artefact

A corrected online guide is also available as a PDF distributed weeks earlier. Draft a distribution inventory: current web page, downloadable PDF, search snippet, archived mirror and printed handout. Identify which items can be replaced and which need an erratum or visible notice. Do not promise retroactive correction of copies held by third parties. Record the residual exposure and a reasonable communication response.

A strong submitted answer should include a one-sentence meaning contract, the observed released behaviour, a distinct hypothesis about the delivery path, a named authorised decision maker and one counterexample that would falsify the proposed correction. A weak answer merely offers a more elegant sentence. When the information is incomplete, the correct professional response is to describe the missing evidence and a safe route for obtaining it, not to guess.

Practice L. Diagnose recurrence after regeneration

A translation patch holds for one day but the same error returns after automated content generation. Compare the source revision, glossary, translation memory match, generation prompt, deployment and cache. Identify whether the reversion came from a stale source, target overwrite or a separate downstream service. Correct the upstream cause where authorised. Test with the next normal regeneration rather than declaring permanent success after a manual hotfix.

A strong submitted answer should include a one-sentence meaning contract, the observed released behaviour, a distinct hypothesis about the delivery path, a named authorised decision maker and one counterexample that would falsify the proposed correction. A weak answer merely offers a more elegant sentence. When the information is incomplete, the correct professional response is to describe the missing evidence and a safe route for obtaining it, not to guess.

Practice M. Separate report identity from language variety

A speaker reports that a translation feels exclusionary. Ask for the exact context, intended reader group and the specific effect, not the reporter’s credentials or a broad claim about their community. Seek domain-appropriate consultation where justified and document which varieties are supported. Separate offence, misunderstanding, style preferences and confirmed unequal access. Do not claim that a single replacement represents all speakers of a language.

A strong submitted answer should include a one-sentence meaning contract, the observed released behaviour, a distinct hypothesis about the delivery path, a named authorised decision maker and one counterexample that would falsify the proposed correction. A weak answer merely offers a more elegant sentence. When the information is incomplete, the correct professional response is to describe the missing evidence and a safe route for obtaining it, not to guess.

Practice N. Make a meaningful closure message

A case was confirmed, corrected and verified on a particular mobile build; the older desktop export remains untested. Write a truthful closure note that states the verified build, the actual correction and the remaining uncertainty. Provide a reopening route. Explain why the phrase ‘fixed everywhere’ would destroy evidence quality. Closure should communicate a bounded conclusion, not erase the limits of the investigation.

A strong submitted answer should include a one-sentence meaning contract, the observed released behaviour, a distinct hypothesis about the delivery path, a named authorised decision maker and one counterexample that would falsify the proposed correction. A weak answer merely offers a more elegant sentence. When the information is incomplete, the correct professional response is to describe the missing evidence and a safe route for obtaining it, not to guess.

Practice O. Design a five-minute training exercise

Give two reviewers the same short source and a deliberately defective target. One reviewer sees only the words; the other sees user state and destination. Compare their decisions, then reveal the meaning contract and release behaviour. Ask each reviewer to revise their judgment and identify which extra context changed it. The point is to teach a reproducible method for recovering intention, not to reward guessing or a preferred prose style.

A strong submitted answer should include a one-sentence meaning contract, the observed released behaviour, a distinct hypothesis about the delivery path, a named authorised decision maker and one counterexample that would falsify the proposed correction. A weak answer merely offers a more elegant sentence. When the information is incomplete, the correct professional response is to describe the missing evidence and a safe route for obtaining it, not to guess.

Practice P. Build a small local correction ledger

Create columns for issue ID, reporting channel, source link, target language, affected content key, released version, consequence, reproduction steps, approved repair, verification evidence and closure. Enter one sample that is semantic and another that is purely technical. Show how to reopen a case when a later release reintroduces the faulty text. Separate personally identifying reporter data from the issue table when it is unnecessary for the investigation.

A strong submitted answer should include a one-sentence meaning contract, the observed released behaviour, a distinct hypothesis about the delivery path, a named authorised decision maker and one counterexample that would falsify the proposed correction. A weak answer merely offers a more elegant sentence. When the information is incomplete, the correct professional response is to describe the missing evidence and a safe route for obtaining it, not to guess.

Practice Q. Interpret quality scores without missing critical cases

A sampled translation quality evaluation finds few errors, but one reported target reverses a prohibition. Explain why an aggregate score cannot prove that this particular released warning is safe. Use the analytic-evaluation lens for type and severity while keeping correction and release verification as separate operational processes. Compare what ISO 5060 says about evaluation with what your organisation must decide about incident containment; do not misstate the standard as prescribing your ticket workflow.

A strong submitted answer should include a one-sentence meaning contract, the observed released behaviour, a distinct hypothesis about the delivery path, a named authorised decision maker and one counterexample that would falsify the proposed correction. A weak answer merely offers a more elegant sentence. When the information is incomplete, the correct professional response is to describe the missing evidence and a safe route for obtaining it, not to guess.

Practice R. Diagnose an overcorrection

A reviewer corrects a culturally awkward greeting by replacing it with a different phrase that changes formal institutional status into friendly endorsement. Compare what each variant signals about authority, relationship and approval. Ask whether the source authorised that change in tone. Produce a third option that is both natural and faithful. Define an acceptance test for readers who need to know whether the notice is official or merely conversational.

A strong submitted answer should include a one-sentence meaning contract, the observed released behaviour, a distinct hypothesis about the delivery path, a named authorised decision maker and one counterexample that would falsify the proposed correction. A weak answer merely offers a more elegant sentence. When the information is incomplete, the correct professional response is to describe the missing evidence and a safe route for obtaining it, not to guess.

Practice S. Use evidence when there is no reproducible failure

A user reports an incorrect target phrase, but the live version, current archived build and source all appear correct. Preserve the report and describe exactly which environments were checked. Seek the reported time, older build or offline surface if available. Do not invent a root cause, and do not imply the user was mistaken. Close as unconfirmed with a reopening path only after a reasonable investigation proportional to the reported risk.

A strong submitted answer should include a one-sentence meaning contract, the observed released behaviour, a distinct hypothesis about the delivery path, a named authorised decision maker and one counterexample that would falsify the proposed correction. A weak answer merely offers a more elegant sentence. When the information is incomplete, the correct professional response is to describe the missing evidence and a safe route for obtaining it, not to guess.

Practice T. Prove a correction survived release

Pick any confirmed case from the preceding diagnostic laboratory. Write a release acceptance criterion that names locale, channel, state, source revision, target meaning and adverse example. Have one reviewer perform the normal path and another the counterexample. Capture text, screenshot or transcript as appropriate, and record the deployed version. The issue closes only when the observed output matches the approved meaning contract and the responsible reviewer accepts the evidence.

A strong submitted answer should include a one-sentence meaning contract, the observed released behaviour, a distinct hypothesis about the delivery path, a named authorised decision maker and one counterexample that would falsify the proposed correction. A weak answer merely offers a more elegant sentence. When the information is incomplete, the correct professional response is to describe the missing evidence and a safe route for obtaining it, not to guess.

A field-ready translation error report template

Report ID and status. Assign a stable identifier and record the received date. Where observed. Record URL, file version, video cue, screen identifier or communication channel. Language. Record the displayed language and specific locale only when relevant. Exact text. Preserve the disputed phrase and nearby context, while removing unrelated personal data. Expected meaning. Record what the reporter understood and the source meaning that a responsible owner later confirms. Impact. Describe the reader action that might change because of the wording. Evidence. Name a reproducible state, safe sample or timestamp. Owner. Name the team or role authorised to decide the correction.

Resolution fields. Record confirmed cause, old and new target, source revision, approval, affected distribution channels, release version, verification method, test result, remaining gaps and reporter response. A checklist may help ensure completeness, but use plain prose when a story is complicated. If the issue is not confirmed, record tested conditions and the reopening path. Never label a repair verified merely because the change request was merged.

The reader’s 90-second decision route

If the wording may affect immediate safety, access to rights, a financial transfer or another hard-to-reverse decision, stop relying on the disputed translation and consult the authoritative source or responsible service before acting. Otherwise preserve the exact phrase, identify where it appears, note your language setting, and state what it seems to mean. Submit this evidence through a legitimate feedback route. If you are a content owner, preserve the report, compare source and released target, confirm the intended action, identify affected output channels and use a governed correction. After deployment, test the same route in the same language and state. The process is complete only when the released result works.

Why a translation correction system needs a human destination

A feedback button without an accountable recipient is a decorative escape hatch. A machine-generated quality score without a source owner is an unsupported opinion. A language expert without access to the actual release state can improve a sentence yet leave the operational defect untouched. A support worker who has no approved correction path can acknowledge the problem but cannot responsibly repair a policy promise. The operational design must therefore join human authority to content identity, technical delivery and reader impact.

That is also why the workflow must permit disagreement and uncertainty. Not every complaint becomes a confirmed error. Not every confirmed error requires the same remedy. Not every published correction reaches every copy. A reliable system makes those distinctions clear so readers are not misled by a superficial show of certainty. Good translation is controlled meaning transfer; good correction is controlled restoration when that transfer has failed.

High-authority sources, publication dates and the limits of these sources

The standards and technical references below are used for their published scopes. The diagnostic workflow and examples in this article are original editorial methods, not verbatim specifications. Availability or scope may change after this article’s research date; consult the official source when applying a standard to a new contractual or regulated context.

ISO 11669:2024, Translation projects — General guidance (published March 2024)

ISO 11669:2024, Translation projects — General guidance (published March 2024). Project specifications, communication among participants, needs analysis, risk assessment and workflow context. This is guidance for translation projects, not a universal mandated incident ticket procedure.

ISO 5060:2024, Translation services — Evaluation of translation output (published February 2024)

ISO 5060:2024, Translation services — Evaluation of translation output (published February 2024). Human analytic evaluation, error types, penalty points, sampling and evaluators. The official abstract states that it does not apply to corrective actions or quality-assurance processes, so this guide keeps evaluation and repair distinct.

ISO 17100:2015, Translation services — Requirements for translation services (2015; ISO confirms current status)

ISO 17100:2015, Translation services — Requirements for translation services (2015; ISO confirms current status). Processes, competences and resources for services meeting applicable specifications; raw machine output plus post-editing and interpreting are outside its stated scope.

ISO 18587:2017, Post-editing of machine translation output (published April 2017; revision under development as of research)

ISO 18587:2017, Post-editing of machine translation output (published April 2017; revision under development as of research). Requirements for full human post-editing of machine translation and post-editor competences. Do not mistake a draft replacement for an already published edition.

W3C WCAG 2.2, Understanding Language of Parts 3.1.2

W3C WCAG 2.2, Understanding Language of Parts 3.1.2. Programmatically identifying language for phrases and passages supports appropriate presentation and pronunciation, especially for assistive technology.

W3C WCAG 2.2, Understanding Error Identification 3.3.1

W3C WCAG 2.2, Understanding Error Identification 3.3.1. Automatically detected form errors should be identified and described in text, useful when a reporter cannot submit a translation complaint.

W3C Internationalization, Language tags in HTML and XML

W3C Internationalization, Language tags in HTML and XML. BCP 47 language tags, language identification and the relationship between page metadata and target-language content.

NIST AI RMF: Generative AI Profile, AI 600-1 (published July 26, 2024)

NIST AI RMF: Generative AI Profile, AI 600-1 (published July 26, 2024). Voluntary cross-sector risk framework for generative AI deployments, relevant to human oversight and evidence-based failure reporting; not a translation quality certificate.

NIST SP 800-61 Revision 3 (published April 2025)

NIST SP 800-61 Revision 3 (published April 2025). Cybersecurity incident response, detection, response and recovery. This article borrows the general discipline of verifiable incident management by analogy; translation errors are not automatically cybersecurity incidents.

Next reading routes without duplicate owners

Return to Master Art of Translation | The Complete System for Moving Meaning Between Languages to revisit meaning, equivalence, source analysis, terminology, translation revision and the complete specialist map. This child owns the post-release feedback-and-correction path: observation, triage, reproduction, authorised repair, deployment verification and closure. It does not compete with the general translation root or earlier specialised quality chapters.

For the release-side accessibility context immediately before this chapter, continue to The Multilingual Web Accessibility System. For learning unfamiliar words, concept families and durable vocabulary retrieval, follow the protected Vocabulary Learning Hub. For grammar, inference, clause structure and writing within English, use the protected How English Works | The Complete Map and Learning Routes. These routes are complementary, not substitute articles.

The next specialist question in this translation architecture is what a team should do when several corrections point to the same upstream cause. The answer begins by distinguishing a single defective sentence from a repeated system fault, then tracing who owns the source, which processes regenerate the error and how a change becomes reliable in every affected destination. Until that investigation is made, a perfectly edited sentence may still be only a temporary illusion of repair.

The last question in any translation correction is not ‘Does the new sentence sound better?’ It is ‘Can we show that the reader now receives the intended meaning under the conditions that previously failed?’