Master Art of Translation | The Spoken Disfluency Fidelity System — How Hesitations, False Starts, Self-Corrections and Interrupted Speech Keep Meaning

How to translate hesitations, false starts and self-corrections accurately

How do you translate a speaker who says “I approved—no, I reviewed the proposal,” or who pauses, repeats, retracts a number, corrects a name, or leaves a sentence unfinished? Translating spoken disfluency means deciding which hesitations, filled pauses, false starts, repetitions, self-corrections and interrupted turns carry meaning, then representing them accurately in another language. The answer depends on whether the result is a verbatim transcript, readable interview translation, documentary subtitles, an accessible caption track, an interpreted exchange or an evidence-sensitive record. Removing every hesitation may make prose smoother while removing the most important part of what happened.

The practical search questions are straightforward: should translators keep “um” and “uh”; how should a false start appear in translation; when can subtitles omit repetitions; how do you translate a speaker correcting “yes” to “no”; how do speech recognition systems distort messy dialogue; and how can a bilingual reviewer check disfluency without making a person sound less competent? This guide gives a repeatable method, annotated examples, decision criteria and quality-control tests. The key is to preserve the speaker’s final commitment and the informative route to it, not mechanically reproduce each source-language sound.

A spontaneous utterance has at least two layers: the proposition the speaker eventually advances, and the visible or audible history of how that proposition emerged. A correction can revoke an earlier proposition; a hesitation can mark planning rather than doubt; an interruption can leave consent unresolved; and a repeated word can be either incidental or deliberately emphatic. Effective speech translation, transcription and subtitling protects these differences. This chapter isolates within-turn disfluency from the neighbouring work of interpersonal clarification and conversation repair, and offers a purpose-specific fidelity system instead of a universal rule to keep or delete fillers.

Quick answers: what to keep, what to simplify, what never to guess

• Always preserve the final number, date, polarity, speaker, commitment, restriction and correction status.
• Keep both the abandoned and repaired phrase when the revision itself matters to evidence, intent, chronology or the story.
• Treat a filler such as “uh” as a potential timing or turn-holding signal; never automatically translate it as fear, deception or uncertainty.
• Make the output brief explicit: evidentiary transcript, edited reading text, timed captions, or interpreting support have different tolerances.
• Use a documented notation for interrupted or uncertain speech, and distinguish inaudible source audio from an unfinished utterance.
• Check accessible captions against audio, important sound cues, speaker identity and timing, not just the polished translation.
• Never turn “I might approve” into “I approved,” or “I approved—sorry, reviewed” into an unqualified approval.
• Review high-risk disfluency with a bilingual human who can consult source audio and the project brief.

Contents

1. The hidden problem: smooth target text can distort speech evidence
2. Mechanisms of disfluency: planning, repair, timing, stance and interruption
3. The eight-layer record and fidelity decision matrix
4. A repeatable twenty-six-pass translation method
5. Failure modes and prevention
6. Worked diagnostics and contrasting outputs
7. Transfer laboratories across translation settings
8. A four-week practice programme
9. Glossary and frequently asked questions
10. Sources, boundaries and next reading routes

The hidden problem: why fluent translation can be wrong

The translator is asked to make speech understandable. That reasonable request can become an implicit instruction to delete the evidence of speech production. The danger is not the removal of a meaningless “um” by itself. It is the unmarked promotion of something the speaker retracted into a settled assertion, the replacement of a provisional term with a more certain term, or the attribution of a completed sentence to a speaker who never reached its final clause. Fluency is one quality dimension; faithful representation of the speaker’s acts is another.

Consider “We sold six—sorry, sixteen units.” A clean output that says “We sold sixteen units” may be appropriate in a concise business summary. A verbatim deposition translation that leaves out the spoken correction is incomplete. A record that says “sixteen, previously six” is helpful if it clearly labels the change as a spoken repair rather than a separate historical sales figure. The same audio supports different legitimate artifacts, but not every transformation is legitimate in every artifact.

Now consider “I, uh, didn’t authorize the release.” Translating the hesitation as if the person was afraid to admit the truth is an interpretation beyond the words. Removing “didn’t” is a catastrophic reversal. A good editor can eliminate the filler in a reading edition, but must preserve the negative commitment. A forensic reviewer may need the timing of the hesitation, and an interpreter may need to convey the denial promptly even while tracking subsequent words.

Spoken language also differs from written prose in syntax. Speakers build clauses incrementally, stop when an audience reacts, revise a noun after seeing an object, and use repetition to keep the floor. Treating every noncanonical sentence as an error to fix imposes written-language expectations on a live activity. The result can misrepresent personality, responsibility or power, particularly when voices differ in accent, age, access needs or communication style.

A system trained to prefer grammatical, fluent output can silently collapse revisions. Modern speech translation and disfluency-correction research recognizes this technical problem, but benchmark findings do not give a single editorial policy for legal records, caption tracks, oral histories and internal summaries. Those policies are human decisions about purpose, risk and permissible loss. A tool can assist alignment and transcription; it cannot declare every abandoned phrase irrelevant.

The decision matrix: different deliverables require different kinds of fidelity

Output	Priority	Basic check
Evidentiary or research transcript	Preserve the observable speech event and allow later analysts to inspect what was spoken, retracted and overlapped.	Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio?
Readable interview translation	Allow readers to follow a person’s answers without mistaking ordinary planning sounds for intentional rhetorical effects.	Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was?
Timed subtitles and accessible captions	Represent spoken content and important audio information under timing and display constraints while maintaining intelligibility.	Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment?
Live interpreting or relay	Transmit a developing utterance in time for another person to respond while maintaining speaker ownership and revision status.	Would the listener take a materially different action because the interpreter concealed the correction or guessed the ending?
Executive summary or edited report	Give the final substantive outcome, not a transcript, with the level of documentation the decision requires.	Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed?

1. Evidentiary or research transcript

Intended use. Preserve the observable speech event and allow later analysts to inspect what was spoken, retracted and overlapped. The first question is what the reader, listener or reviewer must reliably know, not whether the source contains extra syllables. In this setting, the editing threshold follows the information needed for that task, including later challenge or independent checking.

Preserve. Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Review every expression that could change the speaker’s apparent state of knowledge, decision, temporal placement, identity of a participant or responsibility for a claim. If the target language needs different punctuation or a reorganised clause, preserve the relationship even when the exact word order changes.

May simplify. Do not silently remove meaningful abandoned phrases, change polarity, supply missing verbs or collapse timing that bears on disputed meaning. The production note should specify whether the result is verbatim, edited or summarised; do not leave the audience to infer the editorial policy from inconsistent commas and dashes. A bilingual reviewer should be able to compare the source recording with the target and identify the precise point of intervention.

Validation. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? The record may look more hesitant than edited prose, but that is not a defect when reproducibility is the aim. For this reason, fidelity is judged by the legitimate expectations of this output, not a blanket rule that every filler must be visible or that the final grammatical sentence is all that matters.

2. Readable interview translation

Intended use. Allow readers to follow a person’s answers without mistaking ordinary planning sounds for intentional rhetorical effects. The first question is what the reader, listener or reviewer must reliably know, not whether the source contains extra syllables. In this setting, the editing threshold follows the information needed for that task, including later challenge or independent checking.

Preserve. Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Review every expression that could change the speaker’s apparent state of knowledge, decision, temporal placement, identity of a participant or responsibility for a claim. If the target language needs different punctuation or a reorganised clause, preserve the relationship even when the exact word order changes.

May simplify. Routine non-informative fillers may be simplified with disclosure of the editing convention, not with covert changes to commitments. The production note should specify whether the result is verbatim, edited or summarised; do not leave the audience to infer the editorial policy from inconsistent commas and dashes. A bilingual reviewer should be able to compare the source recording with the target and identify the precise point of intervention.

Validation. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? Over-cleaning can erase uncertainty or identity; over-literal reproduction can falsely make a fluent speaker appear incompetent in translation. For this reason, fidelity is judged by the legitimate expectations of this output, not a blanket rule that every filler must be visible or that the final grammatical sentence is all that matters.

3. Timed subtitles and accessible captions

Intended use. Represent spoken content and important audio information under timing and display constraints while maintaining intelligibility. The first question is what the reader, listener or reviewer must reliably know, not whether the source contains extra syllables. In this setting, the editing threshold follows the information needed for that task, including later challenge or independent checking.

Preserve. Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Review every expression that could change the speaker’s apparent state of knowledge, decision, temporal placement, identity of a participant or responsibility for a claim. If the target language needs different punctuation or a reorganised clause, preserve the relationship even when the exact word order changes.

May simplify. Compress expendable repetitions only when the edit does not change emphasis, chronology or consent and does not omit essential dialogue. The production note should specify whether the result is verbatim, edited or summarised; do not leave the audience to infer the editorial policy from inconsistent commas and dashes. A bilingual reviewer should be able to compare the source recording with the target and identify the precise point of intervention.

Validation. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? Excessive words harm readability; excessive compression makes the text contradict or flatten the soundtrack. For this reason, fidelity is judged by the legitimate expectations of this output, not a blanket rule that every filler must be visible or that the final grammatical sentence is all that matters.

4. Live interpreting or relay

Intended use. Transmit a developing utterance in time for another person to respond while maintaining speaker ownership and revision status. The first question is what the reader, listener or reviewer must reliably know, not whether the source contains extra syllables. In this setting, the editing threshold follows the information needed for that task, including later challenge or independent checking.

Preserve. Carry revised numbers, conditions, negatives and identities promptly, explicitly retracting an earlier rendition when necessary. Review every expression that could change the speaker’s apparent state of knowledge, decision, temporal placement, identity of a participant or responsibility for a claim. If the target language needs different punctuation or a reorganised clause, preserve the relationship even when the exact word order changes.

May simplify. Do not pretend a started but unfinished proposition was completed; do not invent coherence to cover overlap or unheard audio. The production note should specify whether the result is verbatim, edited or summarised; do not leave the audience to infer the editorial policy from inconsistent commas and dashes. A bilingual reviewer should be able to compare the source recording with the target and identify the precise point of intervention.

Validation. Would the listener take a materially different action because the interpreter concealed the correction or guessed the ending? A rapid interpretation may commit before the source correction; the remedy is transparent repair, not silent preservation of an error. For this reason, fidelity is judged by the legitimate expectations of this output, not a blanket rule that every filler must be visible or that the final grammatical sentence is all that matters.

5. Executive summary or edited report

Intended use. Give the final substantive outcome, not a transcript, with the level of documentation the decision requires. The first question is what the reader, listener or reviewer must reliably know, not whether the source contains extra syllables. In this setting, the editing threshold follows the information needed for that task, including later challenge or independent checking.

Preserve. Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Review every expression that could change the speaker’s apparent state of knowledge, decision, temporal placement, identity of a participant or responsibility for a claim. If the target language needs different punctuation or a reorganised clause, preserve the relationship even when the exact word order changes.

May simplify. Ordinary speech-production details can often be excluded, provided the document is labelled a summary rather than a literal transcript. The production note should specify whether the result is verbatim, edited or summarised; do not leave the audience to infer the editorial policy from inconsistent commas and dashes. A bilingual reviewer should be able to compare the source recording with the target and identify the precise point of intervention.

Validation. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? Summaries can truthfully compress speech, but they must not dress an inference as a source quotation. For this reason, fidelity is judged by the legitimate expectations of this output, not a blanket rule that every filler must be visible or that the final grammatical sentence is all that matters.

Eight information layers a translator must separate

1. Audio event

A sound happened at a particular time. Record what is heard, distinguish unintelligible sound from deliberate silence, and avoid making a filler into a proposition. For quality assurance, read this layer independently before evaluating fluency. Ask what a careless simplification could remove, what an over-literal copy could invent, and how to annotate unavoidable uncertainty without giving the audience a false impression of precision. It is often better to mark an unresolved point for review than to smooth it into a plausible but unsupported sentence.

2. Speech-production path

The speaker began one form, paused or revised it, and possibly continued. Map the path without automatically deciding that the first form was intended meaning. For quality assurance, read this layer independently before evaluating fluency. Ask what a careless simplification could remove, what an over-literal copy could invent, and how to annotate unavoidable uncertainty without giving the audience a false impression of precision. It is often better to mark an unresolved point for review than to smooth it into a plausible but unsupported sentence.

3. Propositional commitment

Determine the completed assertion, question, instruction or refusal, and whether an earlier version was withdrawn. This layer bears most strongly on material accuracy. For quality assurance, read this layer independently before evaluating fluency. Ask what a careless simplification could remove, what an over-literal copy could invent, and how to annotate unavoidable uncertainty without giving the audience a false impression of precision. It is often better to mark an unresolved point for review than to smooth it into a plausible but unsupported sentence.

4. Epistemic stance

Hedge, confidence, probability and source of knowledge are not the same as ordinary planning hesitation. Preserve lexical markers such as “perhaps” distinctly from “um.” For quality assurance, read this layer independently before evaluating fluency. Ask what a careless simplification could remove, what an over-literal copy could invent, and how to annotate unavoidable uncertainty without giving the audience a false impression of precision. It is often better to mark an unresolved point for review than to smooth it into a plausible but unsupported sentence.

5. Interactional function

A pause may hold the floor, invite a response, yield a turn, signal trouble hearing, or follow an interruption. Context and timing decide; the filler alone does not. For quality assurance, read this layer independently before evaluating fluency. Ask what a careless simplification could remove, what an over-literal copy could invent, and how to annotate unavoidable uncertainty without giving the audience a false impression of precision. It is often better to mark an unresolved point for review than to smooth it into a plausible but unsupported sentence.

6. Voice and identity

Speaker-specific rhythm and code choice may matter to oral history, drama or characterization but should not be converted into crude intelligence or competence judgments. For quality assurance, read this layer independently before evaluating fluency. Ask what a careless simplification could remove, what an over-literal copy could invent, and how to annotate unavoidable uncertainty without giving the audience a false impression of precision. It is often better to mark an unresolved point for review than to smooth it into a plausible but unsupported sentence.

7. Documentary trace

A record may need timecodes, annotation conventions, uncertainty labels, evidence references and an editing log so later readers know what was transformed. For quality assurance, read this layer independently before evaluating fluency. Ask what a careless simplification could remove, what an over-literal copy could invent, and how to annotate unavoidable uncertainty without giving the audience a false impression of precision. It is often better to mark an unresolved point for review than to smooth it into a plausible but unsupported sentence.

8. Target constraints

Captions have screen time and length limits; interpreting has latency; summaries are bounded by purpose. Format decisions must not cancel the first seven layers. For quality assurance, read this layer independently before evaluating fluency. Ask what a careless simplification could remove, what an over-literal copy could invent, and how to annotate unavoidable uncertainty without giving the audience a false impression of precision. It is often better to mark an unresolved point for review than to smooth it into a plausible but unsupported sentence.

Thirty mechanisms of spontaneous speech and their translation consequences

1. Filled pause as planning signal

Recognise the mechanism. A speaker uses “uh” or “um” while searching for wording. The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. The listener has heard a delay, not necessarily a claim of uncertainty or lack of competence. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Do not translate every filler as a lexical hedge; preserve a pause when its timing matters. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

2. Silent pause before a critical negation

Recognise the mechanism. A speaker pauses between an auxiliary and “not.” The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. The location of the pause may influence the listener’s expectation but does not license deletion of the negative. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Check polarity against audio after automated punctuation. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

3. Self-repair of a number

Recognise the mechanism. An initial quantity is replaced by another quantity. The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. The first value can remain visible as a retracted value in a verbatim record, but must not survive as an alternative final fact. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Represent the repair relation, not two independent figures. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

4. Self-repair of a date

Recognise the mechanism. A deadline changes while the speaker is talking. The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. One phrasing may be abandoned immediately; its presence does not establish a prior official deadline. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Preserve the replacement and time qualifier, and separate spoken revision from policy history. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

5. Agent correction

Recognise the mechanism. The subject changes from “we” to “they.” The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. The repair shifts responsibility and cannot be treated as a cosmetic pronoun change. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Reconstruct who does what after the repair. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

6. Object correction

Recognise the mechanism. A speaker replaces the thing that was sent, stored or checked. The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. The object may determine the operational meaning of the instruction. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Keep both referent and correction status when accountability matters. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

7. Modal-strength correction

Recognise the mechanism. “Will” changes to “might,” or “can” to “must.” The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. The speaker retracts a level of commitment or obligation. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Test the translated sentence for force and certainty. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

8. Polarity repair

Recognise the mechanism. “Yes—no” or “do—don’t” appears. The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. The false start and correction may reverse permission or consent. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Never silently keep the first polarity, even when fluent. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

9. Conditional re-entry

Recognise the mechanism. A speaker restarts a clause after “if.” The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. The final condition may govern an entirely different action. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Make the condition scope clear and do not promote the fragment into an unconditional assertion. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

10. Topic restart

Recognise the mechanism. An abandoned topic phrase gives way to another subject. The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. A reader may wrongly interpret two separate subjects or two completed statements. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Distinguish an incomplete launch from a new finished sentence. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

11. Retracing a function word

Recognise the mechanism. “The, the, the” appears before a noun. The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. The repetition often reflects planning rather than emphasis. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Compress where appropriate, but do not remove an intentional contrast elsewhere. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

12. Emphatic repetition

Recognise the mechanism. “No, no, absolutely not” is said with a coherent pragmatic aim. The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. The repetition strengthens refusal rather than documenting word search. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Preserve force or indicate emphasis if the target grammar cannot repeat naturally. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

13. Syllabic false start

Recognise the mechanism. A word begins and is abandoned before completion. The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. The fragment may not have enough lexical identity to translate safely. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Use the project’s cut-off notation rather than guessing the intended word. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

14. Incomplete assertion

Recognise the mechanism. The speaker stops at a point where several completions remain possible. The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. A grammatical target-language sentence might force a claim absent from the audio. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Leave the utterance unfinished or describe uncertainty explicitly. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

15. Interrupted completion

Recognise the mechanism. Another voice cuts off the first speaker. The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. The original speaker’s missing ending cannot be inferred from the interruption alone. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Retain speaker ownership and the truncation. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

16. Overlap and competition for the floor

Recognise the mechanism. Two participants speak simultaneously. The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. Sequential prose can falsely create orderly agreement or a response made after hearing words that were not yet said. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Retain overlap annotation where sequence matters. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

17. Repair with apology

Recognise the mechanism. “Sorry” introduces a changed term. The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. The apology may mark correction rather than culpability. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Translate the repair function without implying a substantive admission of wrongdoing. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

18. Reformulation for the listener

Recognise the mechanism. “That is, what I mean is…” restates an explanation. The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. Both forms may contribute to mutual understanding; the second need not retract the first. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Distinguish elaboration from contradiction. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

19. Pronunciation restart

Recognise the mechanism. A speaker restarts to articulate a technical label. The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. The repair may be phonological only, with no semantic replacement. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Use correct target terminology while preserving the audible restart if documentary fidelity requires it. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

20. Quoted speech correction

Recognise the mechanism. A person revises what they attribute to someone else. The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. The correction affects attribution and should not be collapsed into the translator’s own statement. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Keep quotation boundary and speaker source. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

21. Hedge plus filler

Recognise the mechanism. “Um, probably” combines a filler with an explicit probability claim. The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. The lexical hedge is content-bearing; the filler may be optional depending on brief. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Never delete “probably” while deciding to simplify “um.” Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

22. Laughter during a repair

Recognise the mechanism. A speaker laughs after correcting a mistake. The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. Laughter may qualify affect but does not prove irony, dishonesty or amusement about the topic. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Represent meaningful audio cues cautiously and contextually. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

23. Breath or sigh before an answer

Recognise the mechanism. A nonverbal event precedes a response. The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. A caption may need a sound cue; a clean summary often does not. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Do not translate a sigh as a precise emotion without sufficient context. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

24. Code-switch mid-repair

Recognise the mechanism. A speaker changes language while retrieving a term. The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. The switch may be a search for a name, quotation or word unavailable at the moment. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Retain traceable terminology and avoid inventing sociolinguistic motives. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

25. Dialectal construction versus disfluency

Recognise the mechanism. A grammatical pattern differs from a prescriptive written norm. The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. Not every nonstandard form is an accidental error. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Avoid “repairing” variety into another social identity. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

26. Prosodic boundary ambiguity

Recognise the mechanism. A pause could signal punctuation, hesitation or a new clause. The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. Automatic punctuation can change who did what or how many actions occurred. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Listen for prosody, syntax and later context; document unresolved segmentation. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

27. ASR substitution at restart

Recognise the mechanism. Recognition software confuses the abandoned item with the repaired item. The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. A polished machine output can obscure the actual spoken sequence. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Audit source audio, not only target text. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

28. Timing of a delayed answer

Recognise the mechanism. A response follows a long silence or prompt. The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. Silence may be interactionally relevant but cannot reliably be labelled psychological motive. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Keep duration only if specified by task and supported by recording. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

29. Self-correction after feedback

Recognise the mechanism. A speaker revises wording in response to an audience signal. The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. The change might represent uptake, not a mistaken first draft. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Record the relevant response event when explaining why the repair occurred. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

30. Translation-induced disfluency

Recognise the mechanism. An interpreter repairs their own target wording while conveying the source. The first duty is to inspect the speech sequence rather than immediately replacing it with a tidy sentence. Delays, repairs and restarts can occur at the word, phrase, proposition, turn and interactional levels. A translation that erases their relation may sound elegant while misrepresenting an event that took place in the conversation.

Meaning risk. Source disfluency and interpreter disfluency are separate events. In a careful analysis, identify the speaker, scope, temporal reference, certainty and any material contrast between the abandoned element and the accepted one. The place where the speaker changes course may matter more than the exact surface shape of the filler. A target-language idiom that preserves the relevant distinction is preferable to word-by-word mimicry that creates a new unintended effect.

Operational rule. Do not attribute the interpreter’s error or hesitation to the source speaker. Before sign-off, produce a short audit record that distinguishes what the recording proves from what an editor inferred. Compare the target to the unedited audio at the point of repair, then run a reverse question: would a reader make a materially different decision if this disfluency were missing? If yes, retain or explicitly annotate the information. If no, a properly labelled cleaned edition may simplify it.

The repeatable twenty-six-pass workflow

Pass 1: Obtain the brief

Record target audience, format, legal or research sensitivity, runtime length and whether the output is verbatim, readable or summarised. This step is deliberately separate from writing a polished sentence, because irreversible loss can occur at the earliest cleanup stage. A translator should make the intermediate artifact inspectable: note a timecode, a phrase span, a speaker identifier, an editor decision and a reason that the client can understand. For ordinary low-risk speech, the note can be brief; for disputed content, the evidence must be detailed enough to permit a second reader to reach a different conclusion without reconstructing the entire project.

The practical completion test for pass 1 asks whether the source evidence and target artifact still agree about the speaker’s final commitment. If they do not, correct the target before the next pass. If a decision depends on missing audio, unresolved grammar or a role boundary, label it as unresolved instead of inserting the most likely completion. The repeatable standard is traceable accuracy, proportionate intervention and natural target-language expression, not the mere presence of every spoken syllable.

Pass 2: Secure the authorized source

Identify the correct recording, version, channel, speaker roster and any provided transcript; avoid mixing revised audio files. This step is deliberately separate from writing a polished sentence, because irreversible loss can occur at the earliest cleanup stage. A translator should make the intermediate artifact inspectable: note a timecode, a phrase span, a speaker identifier, an editor decision and a reason that the client can understand. For ordinary low-risk speech, the note can be brief; for disputed content, the evidence must be detailed enough to permit a second reader to reach a different conclusion without reconstructing the entire project.

The practical completion test for pass 2 asks whether the source evidence and target artifact still agree about the speaker’s final commitment. If they do not, correct the target before the next pass. If a decision depends on missing audio, unresolved grammar or a role boundary, label it as unresolved instead of inserting the most likely completion. The repeatable standard is traceable accuracy, proportionate intervention and natural target-language expression, not the mere presence of every spoken syllable.

Pass 3: Map speaker turns

Mark who speaks when, including interruptions, overlaps and uncertain voices; do not infer a speaker from content alone. This step is deliberately separate from writing a polished sentence, because irreversible loss can occur at the earliest cleanup stage. A translator should make the intermediate artifact inspectable: note a timecode, a phrase span, a speaker identifier, an editor decision and a reason that the client can understand. For ordinary low-risk speech, the note can be brief; for disputed content, the evidence must be detailed enough to permit a second reader to reach a different conclusion without reconstructing the entire project.

The practical completion test for pass 3 asks whether the source evidence and target artifact still agree about the speaker’s final commitment. If they do not, correct the target before the next pass. If a decision depends on missing audio, unresolved grammar or a role boundary, label it as unresolved instead of inserting the most likely completion. The repeatable standard is traceable accuracy, proportionate intervention and natural target-language expression, not the mere presence of every spoken syllable.

Pass 4: Create a literal listening transcript

Transcribe what is audible before polishing grammar, including important restarts and incomplete turns. This step is deliberately separate from writing a polished sentence, because irreversible loss can occur at the earliest cleanup stage. A translator should make the intermediate artifact inspectable: note a timecode, a phrase span, a speaker identifier, an editor decision and a reason that the client can understand. For ordinary low-risk speech, the note can be brief; for disputed content, the evidence must be detailed enough to permit a second reader to reach a different conclusion without reconstructing the entire project.

The practical completion test for pass 4 asks whether the source evidence and target artifact still agree about the speaker’s final commitment. If they do not, correct the target before the next pass. If a decision depends on missing audio, unresolved grammar or a role boundary, label it as unresolved instead of inserting the most likely completion. The repeatable standard is traceable accuracy, proportionate intervention and natural target-language expression, not the mere presence of every spoken syllable.

Pass 5: Mark inaudibility distinctly

Use a consistent uncertainty convention; an inaudible word is not a self-interrupted phrase and cannot be invented. This step is deliberately separate from writing a polished sentence, because irreversible loss can occur at the earliest cleanup stage. A translator should make the intermediate artifact inspectable: note a timecode, a phrase span, a speaker identifier, an editor decision and a reason that the client can understand. For ordinary low-risk speech, the note can be brief; for disputed content, the evidence must be detailed enough to permit a second reader to reach a different conclusion without reconstructing the entire project.

The practical completion test for pass 5 asks whether the source evidence and target artifact still agree about the speaker’s final commitment. If they do not, correct the target before the next pass. If a decision depends on missing audio, unresolved grammar or a role boundary, label it as unresolved instead of inserting the most likely completion. The repeatable standard is traceable accuracy, proportionate intervention and natural target-language expression, not the mere presence of every spoken syllable.

Pass 6: Identify fillers

Classify filled pauses, delayed responses and discourse markers by their local function, without imposing motive. This step is deliberately separate from writing a polished sentence, because irreversible loss can occur at the earliest cleanup stage. A translator should make the intermediate artifact inspectable: note a timecode, a phrase span, a speaker identifier, an editor decision and a reason that the client can understand. For ordinary low-risk speech, the note can be brief; for disputed content, the evidence must be detailed enough to permit a second reader to reach a different conclusion without reconstructing the entire project.

The practical completion test for pass 6 asks whether the source evidence and target artifact still agree about the speaker’s final commitment. If they do not, correct the target before the next pass. If a decision depends on missing audio, unresolved grammar or a role boundary, label it as unresolved instead of inserting the most likely completion. The repeatable standard is traceable accuracy, proportionate intervention and natural target-language expression, not the mere presence of every spoken syllable.

Pass 7: Locate the reparandum

Mark the stretch abandoned or superseded by the speaker; distinguish it from surrounding unchanged words. This step is deliberately separate from writing a polished sentence, because irreversible loss can occur at the earliest cleanup stage. A translator should make the intermediate artifact inspectable: note a timecode, a phrase span, a speaker identifier, an editor decision and a reason that the client can understand. For ordinary low-risk speech, the note can be brief; for disputed content, the evidence must be detailed enough to permit a second reader to reach a different conclusion without reconstructing the entire project.

The practical completion test for pass 7 asks whether the source evidence and target artifact still agree about the speaker’s final commitment. If they do not, correct the target before the next pass. If a decision depends on missing audio, unresolved grammar or a role boundary, label it as unresolved instead of inserting the most likely completion. The repeatable standard is traceable accuracy, proportionate intervention and natural target-language expression, not the mere presence of every spoken syllable.

Pass 8: Mark the repair boundary

Find the point where the speaker stops, inserts an editing signal or resumes with a replacement. This step is deliberately separate from writing a polished sentence, because irreversible loss can occur at the earliest cleanup stage. A translator should make the intermediate artifact inspectable: note a timecode, a phrase span, a speaker identifier, an editor decision and a reason that the client can understand. For ordinary low-risk speech, the note can be brief; for disputed content, the evidence must be detailed enough to permit a second reader to reach a different conclusion without reconstructing the entire project.

The practical completion test for pass 8 asks whether the source evidence and target artifact still agree about the speaker’s final commitment. If they do not, correct the target before the next pass. If a decision depends on missing audio, unresolved grammar or a role boundary, label it as unresolved instead of inserting the most likely completion. The repeatable standard is traceable accuracy, proportionate intervention and natural target-language expression, not the mere presence of every spoken syllable.

Pass 9: Identify the repair product

Record the completed replacement and whether it actually supersedes the abandoned stretch or merely elaborates. This step is deliberately separate from writing a polished sentence, because irreversible loss can occur at the earliest cleanup stage. A translator should make the intermediate artifact inspectable: note a timecode, a phrase span, a speaker identifier, an editor decision and a reason that the client can understand. For ordinary low-risk speech, the note can be brief; for disputed content, the evidence must be detailed enough to permit a second reader to reach a different conclusion without reconstructing the entire project.

The practical completion test for pass 9 asks whether the source evidence and target artifact still agree about the speaker’s final commitment. If they do not, correct the target before the next pass. If a decision depends on missing audio, unresolved grammar or a role boundary, label it as unresolved instead of inserting the most likely completion. The repeatable standard is traceable accuracy, proportionate intervention and natural target-language expression, not the mere presence of every spoken syllable.

Pass 10: Protect negation

Check every ‘not’, reversal, rejection and permission formula against the audio before any formatting change. This step is deliberately separate from writing a polished sentence, because irreversible loss can occur at the earliest cleanup stage. A translator should make the intermediate artifact inspectable: note a timecode, a phrase span, a speaker identifier, an editor decision and a reason that the client can understand. For ordinary low-risk speech, the note can be brief; for disputed content, the evidence must be detailed enough to permit a second reader to reach a different conclusion without reconstructing the entire project.

The practical completion test for pass 10 asks whether the source evidence and target artifact still agree about the speaker’s final commitment. If they do not, correct the target before the next pass. If a decision depends on missing audio, unresolved grammar or a role boundary, label it as unresolved instead of inserting the most likely completion. The repeatable standard is traceable accuracy, proportionate intervention and natural target-language expression, not the mere presence of every spoken syllable.

Pass 11: Protect quantities and dates

Verify names, units, decimal markers, deadlines and ranges; note whether the speaker revised them mid-utterance. This step is deliberately separate from writing a polished sentence, because irreversible loss can occur at the earliest cleanup stage. A translator should make the intermediate artifact inspectable: note a timecode, a phrase span, a speaker identifier, an editor decision and a reason that the client can understand. For ordinary low-risk speech, the note can be brief; for disputed content, the evidence must be detailed enough to permit a second reader to reach a different conclusion without reconstructing the entire project.

The practical completion test for pass 11 asks whether the source evidence and target artifact still agree about the speaker’s final commitment. If they do not, correct the target before the next pass. If a decision depends on missing audio, unresolved grammar or a role boundary, label it as unresolved instead of inserting the most likely completion. The repeatable standard is traceable accuracy, proportionate intervention and natural target-language expression, not the mere presence of every spoken syllable.

Pass 12: Protect modality

Check ‘might’, ‘must’, ‘would’, ‘should’, ‘may’ and lexical expressions of confidence separately from filler sounds. This step is deliberately separate from writing a polished sentence, because irreversible loss can occur at the earliest cleanup stage. A translator should make the intermediate artifact inspectable: note a timecode, a phrase span, a speaker identifier, an editor decision and a reason that the client can understand. For ordinary low-risk speech, the note can be brief; for disputed content, the evidence must be detailed enough to permit a second reader to reach a different conclusion without reconstructing the entire project.

The practical completion test for pass 12 asks whether the source evidence and target artifact still agree about the speaker’s final commitment. If they do not, correct the target before the next pass. If a decision depends on missing audio, unresolved grammar or a role boundary, label it as unresolved instead of inserting the most likely completion. The repeatable standard is traceable accuracy, proportionate intervention and natural target-language expression, not the mere presence of every spoken syllable.

Pass 13: Protect attribution

Determine whether a claim belongs to the current speaker, a quoted third party, the interviewer or an interpreter. This step is deliberately separate from writing a polished sentence, because irreversible loss can occur at the earliest cleanup stage. A translator should make the intermediate artifact inspectable: note a timecode, a phrase span, a speaker identifier, an editor decision and a reason that the client can understand. For ordinary low-risk speech, the note can be brief; for disputed content, the evidence must be detailed enough to permit a second reader to reach a different conclusion without reconstructing the entire project.

The practical completion test for pass 13 asks whether the source evidence and target artifact still agree about the speaker’s final commitment. If they do not, correct the target before the next pass. If a decision depends on missing audio, unresolved grammar or a role boundary, label it as unresolved instead of inserting the most likely completion. The repeatable standard is traceable accuracy, proportionate intervention and natural target-language expression, not the mere presence of every spoken syllable.

Pass 14: Resolve reference carefully

Trace pronouns and deictic items only to the extent established by the recording and preceding discourse. This step is deliberately separate from writing a polished sentence, because irreversible loss can occur at the earliest cleanup stage. A translator should make the intermediate artifact inspectable: note a timecode, a phrase span, a speaker identifier, an editor decision and a reason that the client can understand. For ordinary low-risk speech, the note can be brief; for disputed content, the evidence must be detailed enough to permit a second reader to reach a different conclusion without reconstructing the entire project.

The practical completion test for pass 14 asks whether the source evidence and target artifact still agree about the speaker’s final commitment. If they do not, correct the target before the next pass. If a decision depends on missing audio, unresolved grammar or a role boundary, label it as unresolved instead of inserting the most likely completion. The repeatable standard is traceable accuracy, proportionate intervention and natural target-language expression, not the mere presence of every spoken syllable.

Pass 15: Segment by intended output

For caption tracks consider timing; for readable prose consider paragraph sequence; for evidence preserve recoverability. This step is deliberately separate from writing a polished sentence, because irreversible loss can occur at the earliest cleanup stage. A translator should make the intermediate artifact inspectable: note a timecode, a phrase span, a speaker identifier, an editor decision and a reason that the client can understand. For ordinary low-risk speech, the note can be brief; for disputed content, the evidence must be detailed enough to permit a second reader to reach a different conclusion without reconstructing the entire project.

The practical completion test for pass 15 asks whether the source evidence and target artifact still agree about the speaker’s final commitment. If they do not, correct the target before the next pass. If a decision depends on missing audio, unresolved grammar or a role boundary, label it as unresolved instead of inserting the most likely completion. The repeatable standard is traceable accuracy, proportionate intervention and natural target-language expression, not the mere presence of every spoken syllable.

Pass 16: Draft the semantic backbone

Write a target-language representation of the completed, unretracted proposition without silently resolving unanswered questions. This step is deliberately separate from writing a polished sentence, because irreversible loss can occur at the earliest cleanup stage. A translator should make the intermediate artifact inspectable: note a timecode, a phrase span, a speaker identifier, an editor decision and a reason that the client can understand. For ordinary low-risk speech, the note can be brief; for disputed content, the evidence must be detailed enough to permit a second reader to reach a different conclusion without reconstructing the entire project.

The practical completion test for pass 16 asks whether the source evidence and target artifact still agree about the speaker’s final commitment. If they do not, correct the target before the next pass. If a decision depends on missing audio, unresolved grammar or a role boundary, label it as unresolved instead of inserting the most likely completion. The repeatable standard is traceable accuracy, proportionate intervention and natural target-language expression, not the mere presence of every spoken syllable.

Pass 17: Restore consequential disfluency

Reinsert repairs and hesitation signals required for accountability, voice, accessibility or interactional understanding. This step is deliberately separate from writing a polished sentence, because irreversible loss can occur at the earliest cleanup stage. A translator should make the intermediate artifact inspectable: note a timecode, a phrase span, a speaker identifier, an editor decision and a reason that the client can understand. For ordinary low-risk speech, the note can be brief; for disputed content, the evidence must be detailed enough to permit a second reader to reach a different conclusion without reconstructing the entire project.

The practical completion test for pass 17 asks whether the source evidence and target artifact still agree about the speaker’s final commitment. If they do not, correct the target before the next pass. If a decision depends on missing audio, unresolved grammar or a role boundary, label it as unresolved instead of inserting the most likely completion. The repeatable standard is traceable accuracy, proportionate intervention and natural target-language expression, not the mere presence of every spoken syllable.

Pass 18: Choose natural target conventions

Use idiomatic dashes, ellipses, parentheticals or approved notation without exaggerating disorder or certainty. This step is deliberately separate from writing a polished sentence, because irreversible loss can occur at the earliest cleanup stage. A translator should make the intermediate artifact inspectable: note a timecode, a phrase span, a speaker identifier, an editor decision and a reason that the client can understand. For ordinary low-risk speech, the note can be brief; for disputed content, the evidence must be detailed enough to permit a second reader to reach a different conclusion without reconstructing the entire project.

The practical completion test for pass 18 asks whether the source evidence and target artifact still agree about the speaker’s final commitment. If they do not, correct the target before the next pass. If a decision depends on missing audio, unresolved grammar or a role boundary, label it as unresolved instead of inserting the most likely completion. The repeatable standard is traceable accuracy, proportionate intervention and natural target-language expression, not the mere presence of every spoken syllable.

Pass 19: Check timing and speaker labels

For synchronized text, align correction with when the listener would hear it; avoid impossible early disclosure. This step is deliberately separate from writing a polished sentence, because irreversible loss can occur at the earliest cleanup stage. A translator should make the intermediate artifact inspectable: note a timecode, a phrase span, a speaker identifier, an editor decision and a reason that the client can understand. For ordinary low-risk speech, the note can be brief; for disputed content, the evidence must be detailed enough to permit a second reader to reach a different conclusion without reconstructing the entire project.

The practical completion test for pass 19 asks whether the source evidence and target artifact still agree about the speaker’s final commitment. If they do not, correct the target before the next pass. If a decision depends on missing audio, unresolved grammar or a role boundary, label it as unresolved instead of inserting the most likely completion. The repeatable standard is traceable accuracy, proportionate intervention and natural target-language expression, not the mere presence of every spoken syllable.

Pass 20: Run the ‘what changed’ test

Compare propositions before and after the speaker’s repair to document the dimension that changed. This step is deliberately separate from writing a polished sentence, because irreversible loss can occur at the earliest cleanup stage. A translator should make the intermediate artifact inspectable: note a timecode, a phrase span, a speaker identifier, an editor decision and a reason that the client can understand. For ordinary low-risk speech, the note can be brief; for disputed content, the evidence must be detailed enough to permit a second reader to reach a different conclusion without reconstructing the entire project.

The practical completion test for pass 20 asks whether the source evidence and target artifact still agree about the speaker’s final commitment. If they do not, correct the target before the next pass. If a decision depends on missing audio, unresolved grammar or a role boundary, label it as unresolved instead of inserting the most likely completion. The repeatable standard is traceable accuracy, proportionate intervention and natural target-language expression, not the mere presence of every spoken syllable.

Pass 21: Run the ‘what was never said’ test

Identify any target phrase that completes a thought, diagnoses emotion or creates factual certainty absent from the source. This step is deliberately separate from writing a polished sentence, because irreversible loss can occur at the earliest cleanup stage. A translator should make the intermediate artifact inspectable: note a timecode, a phrase span, a speaker identifier, an editor decision and a reason that the client can understand. For ordinary low-risk speech, the note can be brief; for disputed content, the evidence must be detailed enough to permit a second reader to reach a different conclusion without reconstructing the entire project.

The practical completion test for pass 21 asks whether the source evidence and target artifact still agree about the speaker’s final commitment. If they do not, correct the target before the next pass. If a decision depends on missing audio, unresolved grammar or a role boundary, label it as unresolved instead of inserting the most likely completion. The repeatable standard is traceable accuracy, proportionate intervention and natural target-language expression, not the mere presence of every spoken syllable.

Pass 22: Run the register and dignity test

Ensure neutral translation does not stigmatize a speaker for ordinary hesitation, accent or variety. This step is deliberately separate from writing a polished sentence, because irreversible loss can occur at the earliest cleanup stage. A translator should make the intermediate artifact inspectable: note a timecode, a phrase span, a speaker identifier, an editor decision and a reason that the client can understand. For ordinary low-risk speech, the note can be brief; for disputed content, the evidence must be detailed enough to permit a second reader to reach a different conclusion without reconstructing the entire project.

The practical completion test for pass 22 asks whether the source evidence and target artifact still agree about the speaker’s final commitment. If they do not, correct the target before the next pass. If a decision depends on missing audio, unresolved grammar or a role boundary, label it as unresolved instead of inserting the most likely completion. The repeatable standard is traceable accuracy, proportionate intervention and natural target-language expression, not the mere presence of every spoken syllable.

Pass 23: Run the high-risk content test

Check obligations, consent, medical or technical instructions, contractual dates, money and safety conditions one more time. This step is deliberately separate from writing a polished sentence, because irreversible loss can occur at the earliest cleanup stage. A translator should make the intermediate artifact inspectable: note a timecode, a phrase span, a speaker identifier, an editor decision and a reason that the client can understand. For ordinary low-risk speech, the note can be brief; for disputed content, the evidence must be detailed enough to permit a second reader to reach a different conclusion without reconstructing the entire project.

The practical completion test for pass 23 asks whether the source evidence and target artifact still agree about the speaker’s final commitment. If they do not, correct the target before the next pass. If a decision depends on missing audio, unresolved grammar or a role boundary, label it as unresolved instead of inserting the most likely completion. The repeatable standard is traceable accuracy, proportionate intervention and natural target-language expression, not the mere presence of every spoken syllable.

Pass 24: Have an independent bilingual listen

Ask a reviewer unfamiliar with the first draft to annotate disputed or changed items directly from audio. This step is deliberately separate from writing a polished sentence, because irreversible loss can occur at the earliest cleanup stage. A translator should make the intermediate artifact inspectable: note a timecode, a phrase span, a speaker identifier, an editor decision and a reason that the client can understand. For ordinary low-risk speech, the note can be brief; for disputed content, the evidence must be detailed enough to permit a second reader to reach a different conclusion without reconstructing the entire project.

The practical completion test for pass 24 asks whether the source evidence and target artifact still agree about the speaker’s final commitment. If they do not, correct the target before the next pass. If a decision depends on missing audio, unresolved grammar or a role boundary, label it as unresolved instead of inserting the most likely completion. The repeatable standard is traceable accuracy, proportionate intervention and natural target-language expression, not the mere presence of every spoken syllable.

Pass 25: Record approved editorial interventions

Label whether fillers were compressed, how pauses were represented and which uncertainty remains unresolved. This step is deliberately separate from writing a polished sentence, because irreversible loss can occur at the earliest cleanup stage. A translator should make the intermediate artifact inspectable: note a timecode, a phrase span, a speaker identifier, an editor decision and a reason that the client can understand. For ordinary low-risk speech, the note can be brief; for disputed content, the evidence must be detailed enough to permit a second reader to reach a different conclusion without reconstructing the entire project.

The practical completion test for pass 25 asks whether the source evidence and target artifact still agree about the speaker’s final commitment. If they do not, correct the target before the next pass. If a decision depends on missing audio, unresolved grammar or a role boundary, label it as unresolved instead of inserting the most likely completion. The repeatable standard is traceable accuracy, proportionate intervention and natural target-language expression, not the mere presence of every spoken syllable.

Pass 26: Release and preserve a revision trail

Save the signed version, source pointer, glossary and corrections so later readers can reconstruct the editorial choice. This step is deliberately separate from writing a polished sentence, because irreversible loss can occur at the earliest cleanup stage. A translator should make the intermediate artifact inspectable: note a timecode, a phrase span, a speaker identifier, an editor decision and a reason that the client can understand. For ordinary low-risk speech, the note can be brief; for disputed content, the evidence must be detailed enough to permit a second reader to reach a different conclusion without reconstructing the entire project.

The practical completion test for pass 26 asks whether the source evidence and target artifact still agree about the speaker’s final commitment. If they do not, correct the target before the next pass. If a decision depends on missing audio, unresolved grammar or a role boundary, label it as unresolved instead of inserting the most likely completion. The repeatable standard is traceable accuracy, proportionate intervention and natural target-language expression, not the mere presence of every spoken syllable.

Thirty failure modes: diagnosis, consequence and remedy

Failure 1: Silent deletion of an important correction

Symptom: The final value appears, but the disputed original statement disappears. This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Retain the correction in evidence-sensitive outputs and disclose summary status elsewhere. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 2: Treating an abandoned figure as a second fact

Symptom: Two numbers appear as if both are valid measurements. This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Use explicit restart notation and keep the final asserted number distinct. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 3: Removing a negation at an ASR edit

Symptom: “I did not release it” becomes “I did release it.” This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Audit all polarity words against audio with a separate search. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 4: Changing a probability into a promise

Symptom: “Perhaps” or “might” vanishes along with “um.” This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Protect lexical modality before pruning nonlexical fillers. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 5: Psychologizing ordinary fillers

Symptom: A translator infers anxiety, deceit or incompetence from “uh.” This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Render only supported functions, and separate interpretation from transcription. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 6: Inventing a syntactic completion

Symptom: The target finishes an unfinished if-clause. This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Leave it unfinished unless another recorded segment supplies the ending. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 7: Mislabeling an inaudible word as a pause

Symptom: A gap caused by noise is represented as voluntary silence. This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Use separate inaudible/no-audio conventions. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 8: Mislabeling a pause as an inaudible word

Symptom: A deliberate silence suggests missing speech. This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Check waveform, context and recording quality before annotation. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 9: Accidental speaker transfer

Symptom: An overlapping correction is assigned to the wrong speaker. This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Use diarization audit and manually listen to the boundary. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 10: Overpunctuation of a restart

Symptom: A fragment becomes an independent declarative sentence. This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Use syntax and prosody together, not punctuation as a guess. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 11: Loss of repair agency

Symptom: “I—she approved” becomes “I approved.” This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Audit subject and participant roles after every pronoun repair. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 12: Loss of repair object

Symptom: “The valve—the gauge” becomes a statement about the valve. This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Inspect corrected noun phrases and technical labels. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 13: Treating every repeated word as filler

Symptom: An emphatic “never, never” is compressed to “never.” This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Interpret repetition in context and preserve meaningful intensity. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 14: Treating every filler as essential

Symptom: A translated transcript is burdened with unnatural pseudo-disfluency. This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Apply deliverable-specific conventions and disclose editing. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 15: Collapsing apology into fault

Symptom: “Sorry, I mean Thursday” becomes an admission of culpability. This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Translate the discourse repair function without importing blame. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 16: False chronology from a repair

Symptom: A spoken date correction is presented as a rescheduled official deadline. This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Distinguish a verbal self-correction from an institutional change. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 17: Stale machine translation after correction

Symptom: Output preserves the first version but misses the speaker’s later retraction. This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Check streaming systems for finalization and repair propagation. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 18: Wrong subtitle cue timing

Symptom: The corrected value appears before it is spoken and changes audience inference. This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Retain temporal alignment with the audible correction. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 19: Caption omission of relevant audio

Symptom: Important laughter or a door alarm disappears. This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Apply accessibility requirements to meaningful sound as well as words. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 20: Unfair characterization of accent

Symptom: Normal dialect grammar is transcribed as stuttering or broken language. This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Distinguish variety, pronunciation and disfluency. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 21: Compensating by adding mock hesitation

Symptom: A fluent speaker is rendered with invented “uh” sounds. This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Do not manufacture target disfluency for stylistic symmetry. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 22: Missing the scope of ‘only’

Symptom: A restart shifts exclusivity to the wrong noun. This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Rebuild logical scope explicitly before formatting. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 23: Unmarked clean quotation

Symptom: Edited speech is presented inside quotation marks as verbatim. This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Label the cleaned text or retain material utterance details. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 24: Inconsistent pause duration notation

Symptom: A two-second silence is treated as equal to a ten-second gap. This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Use explicit measured timing only when needed and available. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 25: Conflating interviewer prompt and answer

Symptom: A trailing interviewer question becomes part of the respondent’s speech. This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Preserve turn boundaries. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 26: Interpreting speaker emotion as certainty

Symptom: A nervous tone causes the translator to downgrade a clear denial. This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Judge stated commitment independently from delivery. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 27: Ambiguous truncation punctuation

Symptom: A dash is used interchangeably for cut-off, pause and interruption. This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Define project punctuation conventions. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 28: Lost quoted-source boundary

Symptom: A corrected quote becomes the narrator’s direct belief. This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Maintain attribution and quotation scope. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 29: Acronym normalization without evidence

Symptom: A false start near a technical abbreviation is silently expanded into a guess. This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Verify technical term against audio and reference data. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Failure 30: No reproducible edit record

Symptom: Later reviewers cannot tell why a hesitation was omitted. This error frequently passes a fluency-only review because the output reads as fluent prose. The relevant question is whether the speech event has been replaced by a different claim or relationship. Read the source and target aloud in sequence, checking where a revision entered, which words were superseded, whether the speaker reached completion and whether any unstated motive has been attributed to that speaker.

Repair: Save version, brief, timecode and change rationale. Run an adversarial interpretation test: imagine the target is the only document a reader sees. Could that reader infer a stronger statement, an approved action, a different agent or an entirely completed sentence? If so, retain or annotate the missing information. Record the change at the smallest relevant time span so the correction is auditable, and include a second-person review for high-stakes material. A sound translation should not have to choose between a readable surface and an accurate account of who said what.

Worked diagnostics: forty original speech-repair cases

The following scenarios are original teaching examples, not quotations from real interviews, legal evidence, medical records or emergency calls. Their wording uses English to make the mechanism visible without pretending that a literal English rendering is suitable for every target language. For actual bilingual work, the translator must apply the target language’s own grammar, discourse conventions and subtitle or transcript constraints. Each diagnostic deliberately separates source evidence, unsafe simplification, defensible edited meaning and loss-risk reasoning.

Diagnostic 1: Release authorization

Source utterance: “I signed the, uh, I checked the release form. I didn’t sign it.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “I signed the release form.” Safer meaning-first output: “I checked the release form; I did not sign it.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: The verb changes legal agency; the negative resolves the abandoned signing phrase. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: The original speaker did not claim to sign. A polished target that does is a reversal. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “release authorization” case, the protected semantic point is this: The verb changes legal agency; the negative resolves the abandoned signing phrase. The editor should check the exact source phrase “I signed the, uh, I checked the release form. I didn’t sign it.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “I signed the release form.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “release authorization” case, the protected semantic point is this: The verb changes legal agency; the negative resolves the abandoned signing phrase. The editor should check the exact source phrase “I signed the, uh, I checked the release form. I didn’t sign it.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “I signed the release form.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “release authorization” case, the protected semantic point is this: The verb changes legal agency; the negative resolves the abandoned signing phrase. The editor should check the exact source phrase “I signed the, uh, I checked the release form. I didn’t sign it.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “I signed the release form.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “release authorization” case, the protected semantic point is this: The verb changes legal agency; the negative resolves the abandoned signing phrase. The editor should check the exact source phrase “I signed the, uh, I checked the release form. I didn’t sign it.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “I signed the release form.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 2: Inventory revision

Source utterance: “We have forty—sorry, fourteen boxes ready.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “Forty boxes are ready.” Safer meaning-first output: “Fourteen boxes are ready; the speaker initially said ‘forty’ and corrected it.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: The quantity repair matters because the difference could change logistics. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: The first figure is a spoken false start, not a second stock count. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “inventory revision” case, the protected semantic point is this: The quantity repair matters because the difference could change logistics. The editor should check the exact source phrase “We have forty—sorry, fourteen boxes ready.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “Forty boxes are ready.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “inventory revision” case, the protected semantic point is this: The quantity repair matters because the difference could change logistics. The editor should check the exact source phrase “We have forty—sorry, fourteen boxes ready.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “Forty boxes are ready.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “inventory revision” case, the protected semantic point is this: The quantity repair matters because the difference could change logistics. The editor should check the exact source phrase “We have forty—sorry, fourteen boxes ready.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “Forty boxes are ready.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “inventory revision” case, the protected semantic point is this: The quantity repair matters because the difference could change logistics. The editor should check the exact source phrase “We have forty—sorry, fourteen boxes ready.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “Forty boxes are ready.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 3: Deadline false start

Source utterance: “Return it by Tues—by Thursday, assuming the audit closes.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “Return it by Tuesday.” Safer meaning-first output: “Return it by Thursday if the audit closes.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: The revised date is conditional, and the clipped first weekday is not an adopted deadline. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: Keep both qualification and corrected weekday. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “deadline false start” case, the protected semantic point is this: The revised date is conditional, and the clipped first weekday is not an adopted deadline. The editor should check the exact source phrase “Return it by Tues—by Thursday, assuming the audit closes.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “Return it by Tuesday.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “deadline false start” case, the protected semantic point is this: The revised date is conditional, and the clipped first weekday is not an adopted deadline. The editor should check the exact source phrase “Return it by Tues—by Thursday, assuming the audit closes.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “Return it by Tuesday.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “deadline false start” case, the protected semantic point is this: The revised date is conditional, and the clipped first weekday is not an adopted deadline. The editor should check the exact source phrase “Return it by Tues—by Thursday, assuming the audit closes.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “Return it by Tuesday.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “deadline false start” case, the protected semantic point is this: The revised date is conditional, and the clipped first weekday is not an adopted deadline. The editor should check the exact source phrase “Return it by Tues—by Thursday, assuming the audit closes.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “Return it by Tuesday.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 4: Project ownership

Source utterance: “I wrote—we wrote the final draft together.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “I wrote the final draft.” Safer meaning-first output: “We wrote the final draft together.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: Agency changes from singular to collaborative; the final subject matters. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: The source does not establish individual sole authorship. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “project ownership” case, the protected semantic point is this: Agency changes from singular to collaborative; the final subject matters. The editor should check the exact source phrase “I wrote—we wrote the final draft together.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “I wrote the final draft.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “project ownership” case, the protected semantic point is this: Agency changes from singular to collaborative; the final subject matters. The editor should check the exact source phrase “I wrote—we wrote the final draft together.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “I wrote the final draft.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “project ownership” case, the protected semantic point is this: Agency changes from singular to collaborative; the final subject matters. The editor should check the exact source phrase “I wrote—we wrote the final draft together.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “I wrote the final draft.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “project ownership” case, the protected semantic point is this: Agency changes from singular to collaborative; the final subject matters. The editor should check the exact source phrase “I wrote—we wrote the final draft together.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “I wrote the final draft.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 5: Technical component

Source utterance: “Replace the sensor—no, inspect the sensor, replace the cable.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “Replace the sensor and cable.” Safer meaning-first output: “Inspect the sensor; replace the cable.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: An initial imperative is revoked and responsibility separates into two actions. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: Unsafe output could cause avoidable maintenance. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “technical component” case, the protected semantic point is this: An initial imperative is revoked and responsibility separates into two actions. The editor should check the exact source phrase “Replace the sensor—no, inspect the sensor, replace the cable.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “Replace the sensor and cable.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “technical component” case, the protected semantic point is this: An initial imperative is revoked and responsibility separates into two actions. The editor should check the exact source phrase “Replace the sensor—no, inspect the sensor, replace the cable.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “Replace the sensor and cable.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “technical component” case, the protected semantic point is this: An initial imperative is revoked and responsibility separates into two actions. The editor should check the exact source phrase “Replace the sensor—no, inspect the sensor, replace the cable.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “Replace the sensor and cable.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “technical component” case, the protected semantic point is this: An initial imperative is revoked and responsibility separates into two actions. The editor should check the exact source phrase “Replace the sensor—no, inspect the sensor, replace the cable.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “Replace the sensor and cable.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 6: Clinical description

Source utterance: “The patient is stable—well, responsive, but we cannot assess stability yet.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “The patient is stable.” Safer meaning-first output: “The patient responds; stability has not yet been established.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: A diagnostic label is withdrawn in favor of an observable sign. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: Do not promote the initial adjective into a finding. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “clinical description” case, the protected semantic point is this: A diagnostic label is withdrawn in favor of an observable sign. The editor should check the exact source phrase “The patient is stable—well, responsive, but we cannot assess stability yet.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “The patient is stable.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “clinical description” case, the protected semantic point is this: A diagnostic label is withdrawn in favor of an observable sign. The editor should check the exact source phrase “The patient is stable—well, responsive, but we cannot assess stability yet.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “The patient is stable.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “clinical description” case, the protected semantic point is this: A diagnostic label is withdrawn in favor of an observable sign. The editor should check the exact source phrase “The patient is stable—well, responsive, but we cannot assess stability yet.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “The patient is stable.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “clinical description” case, the protected semantic point is this: A diagnostic label is withdrawn in favor of an observable sign. The editor should check the exact source phrase “The patient is stable—well, responsive, but we cannot assess stability yet.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “The patient is stable.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 7: Research estimate

Source utterance: “It is, um, likely around ten, perhaps twelve percent.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “The rate is twelve percent.” Safer meaning-first output: “The speaker estimated roughly ten to twelve percent.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: The number is approximate and a range is being negotiated. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: Do not convert an uncertain estimate into a precise statistic. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “research estimate” case, the protected semantic point is this: The number is approximate and a range is being negotiated. The editor should check the exact source phrase “It is, um, likely around ten, perhaps twelve percent.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “The rate is twelve percent.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “research estimate” case, the protected semantic point is this: The number is approximate and a range is being negotiated. The editor should check the exact source phrase “It is, um, likely around ten, perhaps twelve percent.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “The rate is twelve percent.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “research estimate” case, the protected semantic point is this: The number is approximate and a range is being negotiated. The editor should check the exact source phrase “It is, um, likely around ten, perhaps twelve percent.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “The rate is twelve percent.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “research estimate” case, the protected semantic point is this: The number is approximate and a range is being negotiated. The editor should check the exact source phrase “It is, um, likely around ten, perhaps twelve percent.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “The rate is twelve percent.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 8: Payment approval

Source utterance: “Yes—no. I can review it, but I cannot approve the payment.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “I approved the payment.” Safer meaning-first output: “I can review the payment, not approve it.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: Both polarity and role limit change during the answer. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: A final refusal is not conditional acceptance. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “payment approval” case, the protected semantic point is this: Both polarity and role limit change during the answer. The editor should check the exact source phrase “Yes—no. I can review it, but I cannot approve the payment.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “I approved the payment.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “payment approval” case, the protected semantic point is this: Both polarity and role limit change during the answer. The editor should check the exact source phrase “Yes—no. I can review it, but I cannot approve the payment.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “I approved the payment.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “payment approval” case, the protected semantic point is this: Both polarity and role limit change during the answer. The editor should check the exact source phrase “Yes—no. I can review it, but I cannot approve the payment.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “I approved the payment.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “payment approval” case, the protected semantic point is this: Both polarity and role limit change during the answer. The editor should check the exact source phrase “Yes—no. I can review it, but I cannot approve the payment.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “I approved the payment.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 9: On-camera reaction

Source utterance: “We were delighted—surprised, really, by the result.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “We were delighted.” Safer meaning-first output: “We were surprised by the result.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: The evaluative stance is corrected; the first emotion is retracted or qualified. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: This directly tests the distinction from the prior evaluative stance chapter. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “on-camera reaction” case, the protected semantic point is this: The evaluative stance is corrected; the first emotion is retracted or qualified. The editor should check the exact source phrase “We were delighted—surprised, really, by the result.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “We were delighted.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “on-camera reaction” case, the protected semantic point is this: The evaluative stance is corrected; the first emotion is retracted or qualified. The editor should check the exact source phrase “We were delighted—surprised, really, by the result.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “We were delighted.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “on-camera reaction” case, the protected semantic point is this: The evaluative stance is corrected; the first emotion is retracted or qualified. The editor should check the exact source phrase “We were delighted—surprised, really, by the result.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “We were delighted.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “on-camera reaction” case, the protected semantic point is this: The evaluative stance is corrected; the first emotion is retracted or qualified. The editor should check the exact source phrase “We were delighted—surprised, really, by the result.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “We were delighted.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 10: Minutes dispute

Source utterance: “At the second—actually, third meeting, someone raised the issue.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “The issue was raised at the second meeting.” Safer meaning-first output: “The issue was raised at the third meeting; ‘second’ was corrected.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: Temporal placement is the repaired item, not the fact that an issue was raised. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: Meeting numbers can be important to later verification. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “minutes dispute” case, the protected semantic point is this: Temporal placement is the repaired item, not the fact that an issue was raised. The editor should check the exact source phrase “At the second—actually, third meeting, someone raised the issue.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “The issue was raised at the second meeting.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “minutes dispute” case, the protected semantic point is this: Temporal placement is the repaired item, not the fact that an issue was raised. The editor should check the exact source phrase “At the second—actually, third meeting, someone raised the issue.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “The issue was raised at the second meeting.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “minutes dispute” case, the protected semantic point is this: Temporal placement is the repaired item, not the fact that an issue was raised. The editor should check the exact source phrase “At the second—actually, third meeting, someone raised the issue.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “The issue was raised at the second meeting.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “minutes dispute” case, the protected semantic point is this: Temporal placement is the repaired item, not the fact that an issue was raised. The editor should check the exact source phrase “At the second—actually, third meeting, someone raised the issue.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “The issue was raised at the second meeting.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 11: Travel connection

Source utterance: “We can depart when—if the gate opens.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “Departure happens when the gate opens.” Safer meaning-first output: “Departure is possible only if the gate opens.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: ‘If’ restores contingency; the false start ‘when’ sounds too certain. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: Do not turn a conditional event into an expected event. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “travel connection” case, the protected semantic point is this: ‘If’ restores contingency; the false start ‘when’ sounds too certain. The editor should check the exact source phrase “We can depart when—if the gate opens.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “Departure happens when the gate opens.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “travel connection” case, the protected semantic point is this: ‘If’ restores contingency; the false start ‘when’ sounds too certain. The editor should check the exact source phrase “We can depart when—if the gate opens.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “Departure happens when the gate opens.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “travel connection” case, the protected semantic point is this: ‘If’ restores contingency; the false start ‘when’ sounds too certain. The editor should check the exact source phrase “We can depart when—if the gate opens.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “Departure happens when the gate opens.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “travel connection” case, the protected semantic point is this: ‘If’ restores contingency; the false start ‘when’ sounds too certain. The editor should check the exact source phrase “We can depart when—if the gate opens.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “Departure happens when the gate opens.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 12: Design review

Source utterance: “The prototype failed—it passed the vibration test, but failed thermal testing.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “The prototype failed all tests.” Safer meaning-first output: “It passed vibration testing and failed thermal testing.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: The global verdict is broken into two scoped results. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: Generic ‘failed’ erases a qualified technical outcome. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “design review” case, the protected semantic point is this: The global verdict is broken into two scoped results. The editor should check the exact source phrase “The prototype failed—it passed the vibration test, but failed thermal testing.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “The prototype failed all tests.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “design review” case, the protected semantic point is this: The global verdict is broken into two scoped results. The editor should check the exact source phrase “The prototype failed—it passed the vibration test, but failed thermal testing.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “The prototype failed all tests.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “design review” case, the protected semantic point is this: The global verdict is broken into two scoped results. The editor should check the exact source phrase “The prototype failed—it passed the vibration test, but failed thermal testing.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “The prototype failed all tests.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “design review” case, the protected semantic point is this: The global verdict is broken into two scoped results. The editor should check the exact source phrase “The prototype failed—it passed the vibration test, but failed thermal testing.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “The prototype failed all tests.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 13: Consent discussion

Source utterance: “I agree with the outline, but—not with the final clause.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “I agree with the outline and final clause.” Safer meaning-first output: “I agree with the outline, not the final clause.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: The trailing fragment carries the decisive limitation on agreement. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: Attach the negative to its correct object. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “consent discussion” case, the protected semantic point is this: The trailing fragment carries the decisive limitation on agreement. The editor should check the exact source phrase “I agree with the outline, but—not with the final clause.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “I agree with the outline and final clause.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “consent discussion” case, the protected semantic point is this: The trailing fragment carries the decisive limitation on agreement. The editor should check the exact source phrase “I agree with the outline, but—not with the final clause.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “I agree with the outline and final clause.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “consent discussion” case, the protected semantic point is this: The trailing fragment carries the decisive limitation on agreement. The editor should check the exact source phrase “I agree with the outline, but—not with the final clause.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “I agree with the outline and final clause.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “consent discussion” case, the protected semantic point is this: The trailing fragment carries the decisive limitation on agreement. The editor should check the exact source phrase “I agree with the outline, but—not with the final clause.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “I agree with the outline and final clause.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 14: Witness chronology

Source utterance: “I saw him enter—no, I heard the door, but I did not see anyone.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “I saw him enter.” Safer meaning-first output: “I heard the door; I did not see anyone.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: The evidence source shifts from sight to sound and the person is unverified. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: Do not preserve the abandoned visual claim as testimony. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “witness chronology” case, the protected semantic point is this: The evidence source shifts from sight to sound and the person is unverified. The editor should check the exact source phrase “I saw him enter—no, I heard the door, but I did not see anyone.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “I saw him enter.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “witness chronology” case, the protected semantic point is this: The evidence source shifts from sight to sound and the person is unverified. The editor should check the exact source phrase “I saw him enter—no, I heard the door, but I did not see anyone.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “I saw him enter.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “witness chronology” case, the protected semantic point is this: The evidence source shifts from sight to sound and the person is unverified. The editor should check the exact source phrase “I saw him enter—no, I heard the door, but I did not see anyone.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “I saw him enter.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “witness chronology” case, the protected semantic point is this: The evidence source shifts from sight to sound and the person is unverified. The editor should check the exact source phrase “I saw him enter—no, I heard the door, but I did not see anyone.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “I saw him enter.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 15: Equipment reading

Source utterance: “The gauge shows six point—six point five, not sixty-five.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “The gauge shows sixty-five.” Safer meaning-first output: “The gauge shows 6.5; the speaker explicitly rejects 65.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: The decimal and rejection define a numerically high-risk distinction. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: Check unit and decimal conventions without fabricating extra measurements. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “equipment reading” case, the protected semantic point is this: The decimal and rejection define a numerically high-risk distinction. The editor should check the exact source phrase “The gauge shows six point—six point five, not sixty-five.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “The gauge shows sixty-five.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “equipment reading” case, the protected semantic point is this: The decimal and rejection define a numerically high-risk distinction. The editor should check the exact source phrase “The gauge shows six point—six point five, not sixty-five.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “The gauge shows sixty-five.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “equipment reading” case, the protected semantic point is this: The decimal and rejection define a numerically high-risk distinction. The editor should check the exact source phrase “The gauge shows six point—six point five, not sixty-five.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “The gauge shows sixty-five.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “equipment reading” case, the protected semantic point is this: The decimal and rejection define a numerically high-risk distinction. The editor should check the exact source phrase “The gauge shows six point—six point five, not sixty-five.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “The gauge shows sixty-five.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 16: Office allocation

Source utterance: “She will handle it—could handle it if she has time.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “She will handle it.” Safer meaning-first output: “She could handle it if she has time.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: Promise becomes ability under condition, weakening obligation. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: A report must not allocate a firm task that was not assigned. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “office allocation” case, the protected semantic point is this: Promise becomes ability under condition, weakening obligation. The editor should check the exact source phrase “She will handle it—could handle it if she has time.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “She will handle it.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “office allocation” case, the protected semantic point is this: Promise becomes ability under condition, weakening obligation. The editor should check the exact source phrase “She will handle it—could handle it if she has time.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “She will handle it.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “office allocation” case, the protected semantic point is this: Promise becomes ability under condition, weakening obligation. The editor should check the exact source phrase “She will handle it—could handle it if she has time.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “She will handle it.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “office allocation” case, the protected semantic point is this: Promise becomes ability under condition, weakening obligation. The editor should check the exact source phrase “She will handle it—could handle it if she has time.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “She will handle it.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 17: Training instruction

Source utterance: “Do not switch it on until—I mean, unless the guard is locked.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “Switch it on after some time.” Safer meaning-first output: “Do not switch it on unless the guard is locked.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: ‘Unless’ sets a safety precondition; the first preposition is abandoned. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: The omission could create an unsafe instruction. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “training instruction” case, the protected semantic point is this: ‘Unless’ sets a safety precondition; the first preposition is abandoned. The editor should check the exact source phrase “Do not switch it on until—I mean, unless the guard is locked.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “Switch it on after some time.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “training instruction” case, the protected semantic point is this: ‘Unless’ sets a safety precondition; the first preposition is abandoned. The editor should check the exact source phrase “Do not switch it on until—I mean, unless the guard is locked.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “Switch it on after some time.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “training instruction” case, the protected semantic point is this: ‘Unless’ sets a safety precondition; the first preposition is abandoned. The editor should check the exact source phrase “Do not switch it on until—I mean, unless the guard is locked.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “Switch it on after some time.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “training instruction” case, the protected semantic point is this: ‘Unless’ sets a safety precondition; the first preposition is abandoned. The editor should check the exact source phrase “Do not switch it on until—I mean, unless the guard is locked.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “Switch it on after some time.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 18: Quoted criticism

Source utterance: “He said ‘careless’—or maybe ‘rushed’; I can’t remember.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “He called the work careless.” Safer meaning-first output: “The speaker is unsure whether the other person said ‘careless’ or ‘rushed’.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: The exact quoted adjective is uncertain, and both terms are attributed to another person. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: Do not select one quotation for readability. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “quoted criticism” case, the protected semantic point is this: The exact quoted adjective is uncertain, and both terms are attributed to another person. The editor should check the exact source phrase “He said ‘careless’—or maybe ‘rushed’; I can’t remember.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “He called the work careless.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “quoted criticism” case, the protected semantic point is this: The exact quoted adjective is uncertain, and both terms are attributed to another person. The editor should check the exact source phrase “He said ‘careless’—or maybe ‘rushed’; I can’t remember.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “He called the work careless.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “quoted criticism” case, the protected semantic point is this: The exact quoted adjective is uncertain, and both terms are attributed to another person. The editor should check the exact source phrase “He said ‘careless’—or maybe ‘rushed’; I can’t remember.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “He called the work careless.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “quoted criticism” case, the protected semantic point is this: The exact quoted adjective is uncertain, and both terms are attributed to another person. The editor should check the exact source phrase “He said ‘careless’—or maybe ‘rushed’; I can’t remember.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “He called the work careless.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 19: Apology for name

Source utterance: “Send it to Nora—sorry, to Nola in the records team.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “Send it to Nora.” Safer meaning-first output: “Send it to Nola in the records team.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: A correction of recipient identity is actionable. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: ‘Sorry’ flags a correction and is not an apology for damage. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “apology for name” case, the protected semantic point is this: A correction of recipient identity is actionable. The editor should check the exact source phrase “Send it to Nora—sorry, to Nola in the records team.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “Send it to Nora.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “apology for name” case, the protected semantic point is this: A correction of recipient identity is actionable. The editor should check the exact source phrase “Send it to Nora—sorry, to Nola in the records team.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “Send it to Nora.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “apology for name” case, the protected semantic point is this: A correction of recipient identity is actionable. The editor should check the exact source phrase “Send it to Nora—sorry, to Nola in the records team.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “Send it to Nora.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “apology for name” case, the protected semantic point is this: A correction of recipient identity is actionable. The editor should check the exact source phrase “Send it to Nora—sorry, to Nola in the records team.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “Send it to Nora.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 20: Question restart

Source utterance: “Did you certify—when did you review the file?” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “Did you certify the file?” Safer meaning-first output: “When did you review the file?” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: The speaker shifts from yes-or-no certification to time of review. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: Do not treat the abandoned interrogative as an answered question. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “question restart” case, the protected semantic point is this: The speaker shifts from yes-or-no certification to time of review. The editor should check the exact source phrase “Did you certify—when did you review the file?” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “Did you certify the file?” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “question restart” case, the protected semantic point is this: The speaker shifts from yes-or-no certification to time of review. The editor should check the exact source phrase “Did you certify—when did you review the file?” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “Did you certify the file?” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “question restart” case, the protected semantic point is this: The speaker shifts from yes-or-no certification to time of review. The editor should check the exact source phrase “Did you certify—when did you review the file?” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “Did you certify the file?” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “question restart” case, the protected semantic point is this: The speaker shifts from yes-or-no certification to time of review. The editor should check the exact source phrase “Did you certify—when did you review the file?” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “Did you certify the file?” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 21: Interview childhood

Source utterance: “My father used to—well, my aunt taught me the recipe.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “My father taught me the recipe.” Safer meaning-first output: “My aunt taught me the recipe.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: Family attribution changes at the repair; the first start is incomplete. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: Oral history needs the family relationship not the polished syntax. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “interview childhood” case, the protected semantic point is this: Family attribution changes at the repair; the first start is incomplete. The editor should check the exact source phrase “My father used to—well, my aunt taught me the recipe.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “My father taught me the recipe.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “interview childhood” case, the protected semantic point is this: Family attribution changes at the repair; the first start is incomplete. The editor should check the exact source phrase “My father used to—well, my aunt taught me the recipe.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “My father taught me the recipe.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “interview childhood” case, the protected semantic point is this: Family attribution changes at the repair; the first start is incomplete. The editor should check the exact source phrase “My father used to—well, my aunt taught me the recipe.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “My father taught me the recipe.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “interview childhood” case, the protected semantic point is this: Family attribution changes at the repair; the first start is incomplete. The editor should check the exact source phrase “My father used to—well, my aunt taught me the recipe.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “My father taught me the recipe.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 22: Budget ambiguity

Source utterance: “The total is three—three hundred, excluding tax.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “The total is three.” Safer meaning-first output: “The total is three hundred, excluding tax.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: A restarted number cannot be segmented at the first numeral. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: The exclusion qualifier must stay attached to the corrected total. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “budget ambiguity” case, the protected semantic point is this: A restarted number cannot be segmented at the first numeral. The editor should check the exact source phrase “The total is three—three hundred, excluding tax.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “The total is three.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “budget ambiguity” case, the protected semantic point is this: A restarted number cannot be segmented at the first numeral. The editor should check the exact source phrase “The total is three—three hundred, excluding tax.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “The total is three.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “budget ambiguity” case, the protected semantic point is this: A restarted number cannot be segmented at the first numeral. The editor should check the exact source phrase “The total is three—three hundred, excluding tax.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “The total is three.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “budget ambiguity” case, the protected semantic point is this: A restarted number cannot be segmented at the first numeral. The editor should check the exact source phrase “The total is three—three hundred, excluding tax.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “The total is three.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 23: Emergency radio

Source utterance: “Clear the north—[static]—the east corridor now.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “Clear the north corridor.” Safer meaning-first output: “The audible final instruction refers to the east corridor; earlier words are interrupted by static.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: The missing signal and revised direction require source verification. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: Never guess an unheard direction where safety is involved. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “emergency radio” case, the protected semantic point is this: The missing signal and revised direction require source verification. The editor should check the exact source phrase “Clear the north—[static]—the east corridor now.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “Clear the north corridor.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “emergency radio” case, the protected semantic point is this: The missing signal and revised direction require source verification. The editor should check the exact source phrase “Clear the north—[static]—the east corridor now.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “Clear the north corridor.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “emergency radio” case, the protected semantic point is this: The missing signal and revised direction require source verification. The editor should check the exact source phrase “Clear the north—[static]—the east corridor now.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “Clear the north corridor.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “emergency radio” case, the protected semantic point is this: The missing signal and revised direction require source verification. The editor should check the exact source phrase “Clear the north—[static]—the east corridor now.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “Clear the north corridor.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 24: Public statement

Source utterance: “We guarantee—expect the delivery by Friday.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “We guarantee Friday delivery.” Safer meaning-first output: “We expect delivery by Friday.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: A strong assurance is replaced by a forecast. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: The institution’s commitment changes despite the same date. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “public statement” case, the protected semantic point is this: A strong assurance is replaced by a forecast. The editor should check the exact source phrase “We guarantee—expect the delivery by Friday.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “We guarantee Friday delivery.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “public statement” case, the protected semantic point is this: A strong assurance is replaced by a forecast. The editor should check the exact source phrase “We guarantee—expect the delivery by Friday.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “We guarantee Friday delivery.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “public statement” case, the protected semantic point is this: A strong assurance is replaced by a forecast. The editor should check the exact source phrase “We guarantee—expect the delivery by Friday.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “We guarantee Friday delivery.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “public statement” case, the protected semantic point is this: A strong assurance is replaced by a forecast. The editor should check the exact source phrase “We guarantee—expect the delivery by Friday.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “We guarantee Friday delivery.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 25: Lab notebook dictation

Source utterance: “Sample B—sample D was stored cold.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “Sample B was stored cold.” Safer meaning-first output: “Sample D was stored cold; B was a spoken false start.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: A sample label changes, affecting traceability. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: A reviewer should verify the label with source recording and sample list. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “lab notebook dictation” case, the protected semantic point is this: A sample label changes, affecting traceability. The editor should check the exact source phrase “Sample B—sample D was stored cold.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “Sample B was stored cold.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “lab notebook dictation” case, the protected semantic point is this: A sample label changes, affecting traceability. The editor should check the exact source phrase “Sample B—sample D was stored cold.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “Sample B was stored cold.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “lab notebook dictation” case, the protected semantic point is this: A sample label changes, affecting traceability. The editor should check the exact source phrase “Sample B—sample D was stored cold.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “Sample B was stored cold.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “lab notebook dictation” case, the protected semantic point is this: A sample label changes, affecting traceability. The editor should check the exact source phrase “Sample B—sample D was stored cold.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “Sample B was stored cold.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 26: Call center promise

Source utterance: “Your refund has been—will be considered tomorrow.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “The refund has been issued.” Safer meaning-first output: “The refund will be considered tomorrow.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: The first phrase suggests completed processing; the repair makes future review only. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: Do not confuse consideration with approval or payment. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “call center promise” case, the protected semantic point is this: The first phrase suggests completed processing; the repair makes future review only. The editor should check the exact source phrase “Your refund has been—will be considered tomorrow.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “The refund has been issued.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “call center promise” case, the protected semantic point is this: The first phrase suggests completed processing; the repair makes future review only. The editor should check the exact source phrase “Your refund has been—will be considered tomorrow.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “The refund has been issued.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “call center promise” case, the protected semantic point is this: The first phrase suggests completed processing; the repair makes future review only. The editor should check the exact source phrase “Your refund has been—will be considered tomorrow.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “The refund has been issued.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “call center promise” case, the protected semantic point is this: The first phrase suggests completed processing; the repair makes future review only. The editor should check the exact source phrase “Your refund has been—will be considered tomorrow.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “The refund has been issued.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 27: Audience interruption

Source utterance: “I’ll finish with the last—[audience question]—let me answer that.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “The speaker finished the last point.” Safer meaning-first output: “The speaker was interrupted before completing the last point.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: The source provides no content for the unfinished noun phrase. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: Label the interruption rather than filling the sentence. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “audience interruption” case, the protected semantic point is this: The source provides no content for the unfinished noun phrase. The editor should check the exact source phrase “I’ll finish with the last—[audience question]—let me answer that.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “The speaker finished the last point.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “audience interruption” case, the protected semantic point is this: The source provides no content for the unfinished noun phrase. The editor should check the exact source phrase “I’ll finish with the last—[audience question]—let me answer that.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “The speaker finished the last point.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “audience interruption” case, the protected semantic point is this: The source provides no content for the unfinished noun phrase. The editor should check the exact source phrase “I’ll finish with the last—[audience question]—let me answer that.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “The speaker finished the last point.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “audience interruption” case, the protected semantic point is this: The source provides no content for the unfinished noun phrase. The editor should check the exact source phrase “I’ll finish with the last—[audience question]—let me answer that.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “The speaker finished the last point.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 28: Plan revision

Source utterance: “We should cancel—postpone the workshop until the team returns.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “The workshop was cancelled.” Safer meaning-first output: “The workshop should be postponed until the team returns.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: The action changes from ending an event to deferring it. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: The modal ‘should’ is a recommendation, not an executed decision. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “plan revision” case, the protected semantic point is this: The action changes from ending an event to deferring it. The editor should check the exact source phrase “We should cancel—postpone the workshop until the team returns.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “The workshop was cancelled.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “plan revision” case, the protected semantic point is this: The action changes from ending an event to deferring it. The editor should check the exact source phrase “We should cancel—postpone the workshop until the team returns.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “The workshop was cancelled.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “plan revision” case, the protected semantic point is this: The action changes from ending an event to deferring it. The editor should check the exact source phrase “We should cancel—postpone the workshop until the team returns.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “The workshop was cancelled.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “plan revision” case, the protected semantic point is this: The action changes from ending an event to deferring it. The editor should check the exact source phrase “We should cancel—postpone the workshop until the team returns.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “The workshop was cancelled.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 29: Privacy notice

Source utterance: “We erased—I mean, flagged the records for deletion.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “The records were erased.” Safer meaning-first output: “The records were flagged for deletion.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: The processing stage changes from completed action to pending workflow. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: A legal or compliance text must not claim deletion was completed. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “privacy notice” case, the protected semantic point is this: The processing stage changes from completed action to pending workflow. The editor should check the exact source phrase “We erased—I mean, flagged the records for deletion.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “The records were erased.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “privacy notice” case, the protected semantic point is this: The processing stage changes from completed action to pending workflow. The editor should check the exact source phrase “We erased—I mean, flagged the records for deletion.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “The records were erased.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “privacy notice” case, the protected semantic point is this: The processing stage changes from completed action to pending workflow. The editor should check the exact source phrase “We erased—I mean, flagged the records for deletion.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “The records were erased.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “privacy notice” case, the protected semantic point is this: The processing stage changes from completed action to pending workflow. The editor should check the exact source phrase “We erased—I mean, flagged the records for deletion.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “The records were erased.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 30: Classroom discussion

Source utterance: “The answer is two—wait, no, the variables aren’t independent.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “The answer is two.” Safer meaning-first output: “The speaker withdrew the numerical answer after noticing dependence.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: The repaired epistemic position is unresolved, not a new numerical answer. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: Do not invent a correct result absent from the speech. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “classroom discussion” case, the protected semantic point is this: The repaired epistemic position is unresolved, not a new numerical answer. The editor should check the exact source phrase “The answer is two—wait, no, the variables aren’t independent.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “The answer is two.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “classroom discussion” case, the protected semantic point is this: The repaired epistemic position is unresolved, not a new numerical answer. The editor should check the exact source phrase “The answer is two—wait, no, the variables aren’t independent.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “The answer is two.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “classroom discussion” case, the protected semantic point is this: The repaired epistemic position is unresolved, not a new numerical answer. The editor should check the exact source phrase “The answer is two—wait, no, the variables aren’t independent.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “The answer is two.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “classroom discussion” case, the protected semantic point is this: The repaired epistemic position is unresolved, not a new numerical answer. The editor should check the exact source phrase “The answer is two—wait, no, the variables aren’t independent.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “The answer is two.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 31: Interviewer ambiguity

Source utterance: “Did they stop? Uh, did they stop the machine or the study?” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “Did they stop the machine?” Safer meaning-first output: “The interviewer clarified that the object of ‘stop’ was ambiguous between machine and study.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: The second question disambiguates a prior question rather than answering it. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: Translate both candidates so a respondent can understand the distinction. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “interviewer ambiguity” case, the protected semantic point is this: The second question disambiguates a prior question rather than answering it. The editor should check the exact source phrase “Did they stop? Uh, did they stop the machine or the study?” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “Did they stop the machine?” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “interviewer ambiguity” case, the protected semantic point is this: The second question disambiguates a prior question rather than answering it. The editor should check the exact source phrase “Did they stop? Uh, did they stop the machine or the study?” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “Did they stop the machine?” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “interviewer ambiguity” case, the protected semantic point is this: The second question disambiguates a prior question rather than answering it. The editor should check the exact source phrase “Did they stop? Uh, did they stop the machine or the study?” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “Did they stop the machine?” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “interviewer ambiguity” case, the protected semantic point is this: The second question disambiguates a prior question rather than answering it. The editor should check the exact source phrase “Did they stop? Uh, did they stop the machine or the study?” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “Did they stop the machine?” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 32: Personal recollection

Source utterance: “I never went—well, I visited once, but didn’t stay.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “I never visited.” Safer meaning-first output: “I visited once, but did not stay.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: An absolute claim is corrected into a qualified recollection. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: Do not force the speaker into a contradiction they resolved. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “personal recollection” case, the protected semantic point is this: An absolute claim is corrected into a qualified recollection. The editor should check the exact source phrase “I never went—well, I visited once, but didn’t stay.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “I never visited.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “personal recollection” case, the protected semantic point is this: An absolute claim is corrected into a qualified recollection. The editor should check the exact source phrase “I never went—well, I visited once, but didn’t stay.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “I never visited.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “personal recollection” case, the protected semantic point is this: An absolute claim is corrected into a qualified recollection. The editor should check the exact source phrase “I never went—well, I visited once, but didn’t stay.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “I never visited.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “personal recollection” case, the protected semantic point is this: An absolute claim is corrected into a qualified recollection. The editor should check the exact source phrase “I never went—well, I visited once, but didn’t stay.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “I never visited.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 33: Contract negotiation

Source utterance: “We accept section four—subject to the new audit wording.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “We accept section four without conditions.” Safer meaning-first output: “We accept section four subject to the new audit wording.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: The acceptance is conditional, and the trailing qualification must govern it. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: Punctuation and line breaks must not detach the condition. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “contract negotiation” case, the protected semantic point is this: The acceptance is conditional, and the trailing qualification must govern it. The editor should check the exact source phrase “We accept section four—subject to the new audit wording.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “We accept section four without conditions.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “contract negotiation” case, the protected semantic point is this: The acceptance is conditional, and the trailing qualification must govern it. The editor should check the exact source phrase “We accept section four—subject to the new audit wording.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “We accept section four without conditions.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “contract negotiation” case, the protected semantic point is this: The acceptance is conditional, and the trailing qualification must govern it. The editor should check the exact source phrase “We accept section four—subject to the new audit wording.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “We accept section four without conditions.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “contract negotiation” case, the protected semantic point is this: The acceptance is conditional, and the trailing qualification must govern it. The editor should check the exact source phrase “We accept section four—subject to the new audit wording.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “We accept section four without conditions.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 34: Weather report rehearsal

Source utterance: “It will rain—may rain late in the day.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “Rain is certain.” Safer meaning-first output: “Rain may occur late in the day.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: The weather prediction moves from categorical to probabilistic. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: The replacement modal and time reference are both content-bearing. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “weather report rehearsal” case, the protected semantic point is this: The weather prediction moves from categorical to probabilistic. The editor should check the exact source phrase “It will rain—may rain late in the day.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “Rain is certain.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “weather report rehearsal” case, the protected semantic point is this: The weather prediction moves from categorical to probabilistic. The editor should check the exact source phrase “It will rain—may rain late in the day.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “Rain is certain.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “weather report rehearsal” case, the protected semantic point is this: The weather prediction moves from categorical to probabilistic. The editor should check the exact source phrase “It will rain—may rain late in the day.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “Rain is certain.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “weather report rehearsal” case, the protected semantic point is this: The weather prediction moves from categorical to probabilistic. The editor should check the exact source phrase “It will rain—may rain late in the day.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “Rain is certain.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 35: Software incident

Source utterance: “The server crashed—became unreachable; we don’t know the cause.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “The server crashed.” Safer meaning-first output: “The server became unreachable; the cause remains unknown.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: ‘Crashed’ describes a possible cause not proved by reachability alone. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: Operational reporting should not infer root cause from symptoms. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “software incident” case, the protected semantic point is this: ‘Crashed’ describes a possible cause not proved by reachability alone. The editor should check the exact source phrase “The server crashed—became unreachable; we don’t know the cause.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “The server crashed.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “software incident” case, the protected semantic point is this: ‘Crashed’ describes a possible cause not proved by reachability alone. The editor should check the exact source phrase “The server crashed—became unreachable; we don’t know the cause.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “The server crashed.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “software incident” case, the protected semantic point is this: ‘Crashed’ describes a possible cause not proved by reachability alone. The editor should check the exact source phrase “The server crashed—became unreachable; we don’t know the cause.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “The server crashed.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “software incident” case, the protected semantic point is this: ‘Crashed’ describes a possible cause not proved by reachability alone. The editor should check the exact source phrase “The server crashed—became unreachable; we don’t know the cause.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “The server crashed.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 36: Archive dating

Source utterance: “This was in nineteen eighty—nineteen ninety, I think.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “It was in 1980.” Safer meaning-first output: “The speaker estimated 1990, with explicit uncertainty.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: Both date self-repair and hedge shape the historical claim. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: Do not make tentative memory look like documented chronology. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “archive dating” case, the protected semantic point is this: Both date self-repair and hedge shape the historical claim. The editor should check the exact source phrase “This was in nineteen eighty—nineteen ninety, I think.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “It was in 1980.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “archive dating” case, the protected semantic point is this: Both date self-repair and hedge shape the historical claim. The editor should check the exact source phrase “This was in nineteen eighty—nineteen ninety, I think.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “It was in 1980.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “archive dating” case, the protected semantic point is this: Both date self-repair and hedge shape the historical claim. The editor should check the exact source phrase “This was in nineteen eighty—nineteen ninety, I think.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “It was in 1980.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “archive dating” case, the protected semantic point is this: Both date self-repair and hedge shape the historical claim. The editor should check the exact source phrase “This was in nineteen eighty—nineteen ninety, I think.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “It was in 1980.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 37: Repaired adjective

Source utterance: “The treatment is harmless—well, lower risk under these conditions.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “The treatment is harmless.” Safer meaning-first output: “The speaker describes lower risk under stated conditions.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: A claim of zero harm is replaced with a comparative, bounded claim. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: Avoid unqualified safety assurances. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “repaired adjective” case, the protected semantic point is this: A claim of zero harm is replaced with a comparative, bounded claim. The editor should check the exact source phrase “The treatment is harmless—well, lower risk under these conditions.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “The treatment is harmless.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “repaired adjective” case, the protected semantic point is this: A claim of zero harm is replaced with a comparative, bounded claim. The editor should check the exact source phrase “The treatment is harmless—well, lower risk under these conditions.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “The treatment is harmless.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “repaired adjective” case, the protected semantic point is this: A claim of zero harm is replaced with a comparative, bounded claim. The editor should check the exact source phrase “The treatment is harmless—well, lower risk under these conditions.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “The treatment is harmless.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “repaired adjective” case, the protected semantic point is this: A claim of zero harm is replaced with a comparative, bounded claim. The editor should check the exact source phrase “The treatment is harmless—well, lower risk under these conditions.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “The treatment is harmless.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 38: Conditional approval

Source utterance: “You can publish if—after the reviewer signs off.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “You can publish now.” Safer meaning-first output: “Publication is permitted after reviewer sign-off.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: A false start becomes a temporal gate; action is not immediate. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: Do not confuse permission with fulfilment of prerequisite. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “conditional approval” case, the protected semantic point is this: A false start becomes a temporal gate; action is not immediate. The editor should check the exact source phrase “You can publish if—after the reviewer signs off.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “You can publish now.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “conditional approval” case, the protected semantic point is this: A false start becomes a temporal gate; action is not immediate. The editor should check the exact source phrase “You can publish if—after the reviewer signs off.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “You can publish now.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “conditional approval” case, the protected semantic point is this: A false start becomes a temporal gate; action is not immediate. The editor should check the exact source phrase “You can publish if—after the reviewer signs off.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “You can publish now.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “conditional approval” case, the protected semantic point is this: A false start becomes a temporal gate; action is not immediate. The editor should check the exact source phrase “You can publish if—after the reviewer signs off.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “You can publish now.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 39: Inventory certainty

Source utterance: “We counted, um, maybe forty units, not exactly.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “There are exactly forty.” Safer meaning-first output: “The count is an approximate forty units.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: Hesitation combines with lexical approximators that cannot be removed. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: Do not make a precise ledger statement from an estimate. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “inventory certainty” case, the protected semantic point is this: Hesitation combines with lexical approximators that cannot be removed. The editor should check the exact source phrase “We counted, um, maybe forty units, not exactly.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “There are exactly forty.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “inventory certainty” case, the protected semantic point is this: Hesitation combines with lexical approximators that cannot be removed. The editor should check the exact source phrase “We counted, um, maybe forty units, not exactly.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “There are exactly forty.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “inventory certainty” case, the protected semantic point is this: Hesitation combines with lexical approximators that cannot be removed. The editor should check the exact source phrase “We counted, um, maybe forty units, not exactly.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “There are exactly forty.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “inventory certainty” case, the protected semantic point is this: Hesitation combines with lexical approximators that cannot be removed. The editor should check the exact source phrase “We counted, um, maybe forty units, not exactly.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “There are exactly forty.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Diagnostic 40: Video caption cut-off

Source utterance: “The last step is to—[alarm]—stop and move away.” The initial job is to locate the exact switch in the speaker’s production path. It is not enough to recognise the final clause. Mark the earliest audible commitment, the point of hesitation or replacement, and any remaining qualification. Decide whether the apparent start was a grammatical prediction, an actual assertion, or a segment never finished. This establishes what the translator is entitled to say in a finished target sentence.

Unsafe polished output: “The last step is to stop.” Safer meaning-first output: “An alarm interrupts an unfinished explanation, followed by an instruction to move away.” The safer line is an explanatory English rendering rather than a mandatory script for all languages. In an evidentiary translation, the source’s abandoned material may need to stay visible with agreed punctuation; in a reading version the accepted proposition may stand alone if the publication describes that editing convention. It is the relation between versions, not a particular dash style, that governs fidelity.

Mechanism diagnosis: Meaningful sound and turn interruption affect caption accessibility. Listen again to identify what was actually heard and who uttered it. Do not upgrade an early fragment into a complete statement; do not treat a later correction as proof that a formal policy or external fact changed. Notice whether negation, scope, modality, reference, number or the speaker’s level of knowledge moves across the repair. These variables change practical outcomes even when the resulting target prose still reads naturally.

Adversarial audit: Keep relevant non-speech audio and the final emergency action. The reviewer should ask two independent questions. First, could a reader act on the target and obtain the wrong result or assign responsibility to the wrong person? Second, could a listener comparing the original audio with the edited text explain the omitted restart? If either answer is troubling, preserve the critical repair or add a transparent note. Silence about editorial intervention is not neutrality when the intervention changes the event represented.

Evidentiary or research transcript decision: Keep relevant fillers, repetitions, self-repairs, pauses and uncertainty using agreed marks; preserve time anchors and speaker turns. Applied to the “video caption cut-off” case, the protected semantic point is this: Meaningful sound and turn interruption affect caption accessibility. The editor should check the exact source phrase “The last step is to—[alarm]—stop and move away.” against the proposed target. Can a reviewer reconstruct the contrast between the abandoned and accepted formulation while listening to the audio? A legitimate shortened line must not become “The last step is to stop.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Readable interview translation decision: Keep repairs that revise facts, agency, attitude or narrative sequence; preserve characteristic patterned speech when needed for voice. Applied to the “video caption cut-off” case, the protected semantic point is this: Meaningful sound and turn interruption affect caption accessibility. The editor should check the exact source phrase “The last step is to—[alarm]—stop and move away.” against the proposed target. Does the cleaned passage make any answer more certain, categorical, polished or completed than the source actually was? A legitimate shortened line must not become “The last step is to stop.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Timed subtitles and accessible captions decision: Retain consequential repairs, final proposition and the information required to understand action; make speaker changes and meaningful audio cues clear. Applied to the “video caption cut-off” case, the protected semantic point is this: Meaningful sound and turn interruption affect caption accessibility. The editor should check the exact source phrase “The last step is to—[alarm]—stop and move away.” against the proposed target. Can a viewer who cannot hear the audio understand the same crucial correction, refusal or interruption at approximately the same moment? A legitimate shortened line must not become “The last step is to stop.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Executive summary or edited report decision: Retain corrected facts and qualify decisions not reached; note important revisions or disputes in a provenance note when they explain the outcome. Applied to the “video caption cut-off” case, the protected semantic point is this: Meaningful sound and turn interruption affect caption accessibility. The editor should check the exact source phrase “The last step is to—[alarm]—stop and move away.” against the proposed target. Could a reader mistake the report for a verbatim quotation, or conclude that a proposal was approved when only reviewed? A legitimate shortened line must not become “The last step is to stop.” merely because that wording is shorter. An annotation, one additional caption cue or a brief note may preserve the distinction without burdening every line of the delivery.

Cross-format transfer: how the same case changes without changing its facts

Transfer laboratory 1: The research interview and the reader

Prepare an unedited transcript, a lightly edited translated interview and an interpretive summary. In the first keep meaningful self-repairs and a consistent notation; in the second remove only uninformative fillers while retaining corrected agency and dates; in the third report the final claim but state whether it is an estimate or recollection. Compare the three artifacts against one audio segment and mark all editorial choices. Record the concrete edit decision, the phrase that justifies it and the audience that will receive the result. Then carry out a three-way review with the source audio, an unedited source transcript and the translated deliverable. Do not let a successful style review substitute for evidence that the final commitment, sequence and qualification survive.

For this laboratory, prepare a second output with the opposite editorial purpose, then compare what legitimately changes and what must remain invariant. The target may compress ordinary planning pauses for readable communication, yet it must still protect polarity, agency, dates, quantities, conditionality, quotation ownership and incompleteness. If two reviewers disagree, isolate the recording span and record the unresolved reading rather than voting to hide it. Make a short client-facing rationale describing how the choice balances accuracy, accessibility and readability.

A useful mastery test is to hand only the translated result to a new reader and ask what happened, who committed to what, and what is still unknown. If the answer differs materially from what careful listeners can establish in the source, revise. If the differences are only the predictable differences between a verbatim transcript and a disclosed cleaned edition, document that distinction. This is transfer: the method survives changes of genre because the semantic invariants are identified before surface editing.

Transfer laboratory 2: The live caption and the archived transcript

Work with a recording whose speaker corrects a technical name after a three-second pause. The live caption needs temporal fidelity and must not show the correction prematurely. The archived transcript can mark the abandoned term and repair boundary with a timecode. Verify that both communicate the same adopted term and neither makes the pause into an assertion of ignorance. Record the concrete edit decision, the phrase that justifies it and the audience that will receive the result. Then carry out a three-way review with the source audio, an unedited source transcript and the translated deliverable. Do not let a successful style review substitute for evidence that the final commitment, sequence and qualification survive.

For this laboratory, prepare a second output with the opposite editorial purpose, then compare what legitimately changes and what must remain invariant. The target may compress ordinary planning pauses for readable communication, yet it must still protect polarity, agency, dates, quantities, conditionality, quotation ownership and incompleteness. If two reviewers disagree, isolate the recording span and record the unresolved reading rather than voting to hide it. Make a short client-facing rationale describing how the choice balances accuracy, accessibility and readability.

A useful mastery test is to hand only the translated result to a new reader and ask what happened, who committed to what, and what is still unknown. If the answer differs materially from what careful listeners can establish in the source, revise. If the differences are only the predictable differences between a verbatim transcript and a disclosed cleaned edition, document that distinction. This is transfer: the method survives changes of genre because the semantic invariants are identified before surface editing.

Transfer laboratory 3: The interpreter’s early commitment

Suppose the source begins with a future promise and corrects it to a conditional possibility after the interpreter has already rendered the stronger clause. The ethical response is not to leave the stronger target sentence uncorrected. Rehearse a concise repair formula that clearly attributes the new modal force to the original speaker’s revision, without creating a separate policy announcement. Record the concrete edit decision, the phrase that justifies it and the audience that will receive the result. Then carry out a three-way review with the source audio, an unedited source transcript and the translated deliverable. Do not let a successful style review substitute for evidence that the final commitment, sequence and qualification survive.

For this laboratory, prepare a second output with the opposite editorial purpose, then compare what legitimately changes and what must remain invariant. The target may compress ordinary planning pauses for readable communication, yet it must still protect polarity, agency, dates, quantities, conditionality, quotation ownership and incompleteness. If two reviewers disagree, isolate the recording span and record the unresolved reading rather than voting to hide it. Make a short client-facing rationale describing how the choice balances accuracy, accessibility and readability.

A useful mastery test is to hand only the translated result to a new reader and ask what happened, who committed to what, and what is still unknown. If the answer differs materially from what careful listeners can establish in the source, revise. If the differences are only the predictable differences between a verbatim transcript and a disclosed cleaned edition, document that distinction. This is transfer: the method survives changes of genre because the semantic invariants are identified before surface editing.

Transfer laboratory 4: The machine-generated fluent transcript

Feed an imagined ASR output into a bilingual review process and compare it with the original spoken exchange. Highlight a lost ‘not,’ an abandoned digit, a misheard decimal and a turn boundary error. Each may be nearly invisible to a fluency judge. Create an error log with timestamp, risk score, restoration action and whether the final text is fit for an evidence-sensitive purpose. Record the concrete edit decision, the phrase that justifies it and the audience that will receive the result. Then carry out a three-way review with the source audio, an unedited source transcript and the translated deliverable. Do not let a successful style review substitute for evidence that the final commitment, sequence and qualification survive.

For this laboratory, prepare a second output with the opposite editorial purpose, then compare what legitimately changes and what must remain invariant. The target may compress ordinary planning pauses for readable communication, yet it must still protect polarity, agency, dates, quantities, conditionality, quotation ownership and incompleteness. If two reviewers disagree, isolate the recording span and record the unresolved reading rather than voting to hide it. Make a short client-facing rationale describing how the choice balances accuracy, accessibility and readability.

A useful mastery test is to hand only the translated result to a new reader and ask what happened, who committed to what, and what is still unknown. If the answer differs materially from what careful listeners can establish in the source, revise. If the differences are only the predictable differences between a verbatim transcript and a disclosed cleaned edition, document that distinction. This is transfer: the method survives changes of genre because the semantic invariants are identified before surface editing.

Transfer laboratory 5: The public-facing interview quote

The editor is tempted to quote a beautifully grammatical assertion that only exists after several silent deletions. Compose two alternatives: a faithful quotation with meaningful repairs visible, and a clearly labelled paraphrase preserving the conclusion. Explain why moving a paraphrase into quotation marks changes the reader’s expectations about exact wording. Record the concrete edit decision, the phrase that justifies it and the audience that will receive the result. Then carry out a three-way review with the source audio, an unedited source transcript and the translated deliverable. Do not let a successful style review substitute for evidence that the final commitment, sequence and qualification survive.

For this laboratory, prepare a second output with the opposite editorial purpose, then compare what legitimately changes and what must remain invariant. The target may compress ordinary planning pauses for readable communication, yet it must still protect polarity, agency, dates, quantities, conditionality, quotation ownership and incompleteness. If two reviewers disagree, isolate the recording span and record the unresolved reading rather than voting to hide it. Make a short client-facing rationale describing how the choice balances accuracy, accessibility and readability.

A useful mastery test is to hand only the translated result to a new reader and ask what happened, who committed to what, and what is still unknown. If the answer differs materially from what careful listeners can establish in the source, revise. If the differences are only the predictable differences between a verbatim transcript and a disclosed cleaned edition, document that distinction. This is transfer: the method survives changes of genre because the semantic invariants are identified before surface editing.

Transfer laboratory 6: The oral history with uncertain memory

A speaker corrects a year, pauses and says ‘I think.’ Produce a readable passage that keeps the uncertainty and a transcription note that preserves the first year. Research the event independently only in a separate editorial annotation, because external fact-checking does not license changing what the witness remembers having said. Record the concrete edit decision, the phrase that justifies it and the audience that will receive the result. Then carry out a three-way review with the source audio, an unedited source transcript and the translated deliverable. Do not let a successful style review substitute for evidence that the final commitment, sequence and qualification survive.

For this laboratory, prepare a second output with the opposite editorial purpose, then compare what legitimately changes and what must remain invariant. The target may compress ordinary planning pauses for readable communication, yet it must still protect polarity, agency, dates, quantities, conditionality, quotation ownership and incompleteness. If two reviewers disagree, isolate the recording span and record the unresolved reading rather than voting to hide it. Make a short client-facing rationale describing how the choice balances accuracy, accessibility and readability.

A useful mastery test is to hand only the translated result to a new reader and ask what happened, who committed to what, and what is still unknown. If the answer differs materially from what careful listeners can establish in the source, revise. If the differences are only the predictable differences between a verbatim transcript and a disclosed cleaned edition, document that distinction. This is transfer: the method survives changes of genre because the semantic invariants are identified before surface editing.

Transfer laboratory 7: The safety instruction at a false start

Build a QA checklist for commands that contain ‘not,’ ‘unless,’ ‘before,’ ‘after’ and a repaired equipment name. The release criterion is not grammatical elegance but that the listener will perform the intended action under the intended precondition. Require second-person audio verification before the target instruction is used operationally. Record the concrete edit decision, the phrase that justifies it and the audience that will receive the result. Then carry out a three-way review with the source audio, an unedited source transcript and the translated deliverable. Do not let a successful style review substitute for evidence that the final commitment, sequence and qualification survive.

For this laboratory, prepare a second output with the opposite editorial purpose, then compare what legitimately changes and what must remain invariant. The target may compress ordinary planning pauses for readable communication, yet it must still protect polarity, agency, dates, quantities, conditionality, quotation ownership and incompleteness. If two reviewers disagree, isolate the recording span and record the unresolved reading rather than voting to hide it. Make a short client-facing rationale describing how the choice balances accuracy, accessibility and readability.

A useful mastery test is to hand only the translated result to a new reader and ask what happened, who committed to what, and what is still unknown. If the answer differs materially from what careful listeners can establish in the source, revise. If the differences are only the predictable differences between a verbatim transcript and a disclosed cleaned edition, document that distinction. This is transfer: the method survives changes of genre because the semantic invariants are identified before surface editing.

Transfer laboratory 8: The overlapping meeting

Two speakers say ‘yes’ and ‘no’ at the same time, then one restarts. Produce a timed multi-speaker transcript rather than a single, merged sentence. Annotate which words overlap and avoid inventing a collective decision from simultaneous utterances. Review whether a later summary should say no decision was reached. Record the concrete edit decision, the phrase that justifies it and the audience that will receive the result. Then carry out a three-way review with the source audio, an unedited source transcript and the translated deliverable. Do not let a successful style review substitute for evidence that the final commitment, sequence and qualification survive.

For this laboratory, prepare a second output with the opposite editorial purpose, then compare what legitimately changes and what must remain invariant. The target may compress ordinary planning pauses for readable communication, yet it must still protect polarity, agency, dates, quantities, conditionality, quotation ownership and incompleteness. If two reviewers disagree, isolate the recording span and record the unresolved reading rather than voting to hide it. Make a short client-facing rationale describing how the choice balances accuracy, accessibility and readability.

A useful mastery test is to hand only the translated result to a new reader and ask what happened, who committed to what, and what is still unknown. If the answer differs materially from what careful listeners can establish in the source, revise. If the differences are only the predictable differences between a verbatim transcript and a disclosed cleaned edition, document that distinction. This is transfer: the method survives changes of genre because the semantic invariants are identified before surface editing.

Transfer laboratory 9: The audiobook and dramatized dialogue

A character’s carefully staged repetition and abandoned word may shape tension. Try a literary translation that preserves pacing through target-language phrasing rather than reproducing each syllable. Compare it with a documentary transcription and explain why a more dramatic imitation could distort the character’s voice while a sterile edit removes deliberate craft. Record the concrete edit decision, the phrase that justifies it and the audience that will receive the result. Then carry out a three-way review with the source audio, an unedited source transcript and the translated deliverable. Do not let a successful style review substitute for evidence that the final commitment, sequence and qualification survive.

For this laboratory, prepare a second output with the opposite editorial purpose, then compare what legitimately changes and what must remain invariant. The target may compress ordinary planning pauses for readable communication, yet it must still protect polarity, agency, dates, quantities, conditionality, quotation ownership and incompleteness. If two reviewers disagree, isolate the recording span and record the unresolved reading rather than voting to hide it. Make a short client-facing rationale describing how the choice balances accuracy, accessibility and readability.

A useful mastery test is to hand only the translated result to a new reader and ask what happened, who committed to what, and what is still unknown. If the answer differs materially from what careful listeners can establish in the source, revise. If the differences are only the predictable differences between a verbatim transcript and a disclosed cleaned edition, document that distinction. This is transfer: the method survives changes of genre because the semantic invariants are identified before surface editing.

Transfer laboratory 10: The terminology workshop

Take five repaired technical nouns and build a reference table with audio location, abandoned source label, corrected source label, approved target term and domain definition. Avoid treating pronunciation fragments as evidence for unverified acronyms. Require a specialist to approve term identity before finalizing the sentence around it. Record the concrete edit decision, the phrase that justifies it and the audience that will receive the result. Then carry out a three-way review with the source audio, an unedited source transcript and the translated deliverable. Do not let a successful style review substitute for evidence that the final commitment, sequence and qualification survive.

For this laboratory, prepare a second output with the opposite editorial purpose, then compare what legitimately changes and what must remain invariant. The target may compress ordinary planning pauses for readable communication, yet it must still protect polarity, agency, dates, quantities, conditionality, quotation ownership and incompleteness. If two reviewers disagree, isolate the recording span and record the unresolved reading rather than voting to hide it. Make a short client-facing rationale describing how the choice balances accuracy, accessibility and readability.

A useful mastery test is to hand only the translated result to a new reader and ask what happened, who committed to what, and what is still unknown. If the answer differs materially from what careful listeners can establish in the source, revise. If the differences are only the predictable differences between a verbatim transcript and a disclosed cleaned edition, document that distinction. This is transfer: the method survives changes of genre because the semantic invariants are identified before surface editing.

Transfer laboratory 11: The accessibility review

Watch a segment with speech, laughter, an alarm and a repaired instruction. Draft captions with speaker labels where needed, the important sound cues and the final instruction. Check timing and readability while listening with the captions. A deletion is acceptable only if essential dialogue and audio context still reach the viewer. Record the concrete edit decision, the phrase that justifies it and the audience that will receive the result. Then carry out a three-way review with the source audio, an unedited source transcript and the translated deliverable. Do not let a successful style review substitute for evidence that the final commitment, sequence and qualification survive.

For this laboratory, prepare a second output with the opposite editorial purpose, then compare what legitimately changes and what must remain invariant. The target may compress ordinary planning pauses for readable communication, yet it must still protect polarity, agency, dates, quantities, conditionality, quotation ownership and incompleteness. If two reviewers disagree, isolate the recording span and record the unresolved reading rather than voting to hide it. Make a short client-facing rationale describing how the choice balances accuracy, accessibility and readability.

A useful mastery test is to hand only the translated result to a new reader and ask what happened, who committed to what, and what is still unknown. If the answer differs materially from what careful listeners can establish in the source, revise. If the differences are only the predictable differences between a verbatim transcript and a disclosed cleaned edition, document that distinction. This is transfer: the method survives changes of genre because the semantic invariants are identified before surface editing.

Transfer laboratory 12: The speaker-fairness audit

Compare two translations of equally ordinary pauses from speakers with different accents. Ask whether the editor labels one as incoherent while smoothly normalizing the other. Separate language variation from disfluency, and distinguish a documentary choice to record a filler from an editorial opinion about competence. Rewrite the target to preserve dignity and factual meaning. Record the concrete edit decision, the phrase that justifies it and the audience that will receive the result. Then carry out a three-way review with the source audio, an unedited source transcript and the translated deliverable. Do not let a successful style review substitute for evidence that the final commitment, sequence and qualification survive.

For this laboratory, prepare a second output with the opposite editorial purpose, then compare what legitimately changes and what must remain invariant. The target may compress ordinary planning pauses for readable communication, yet it must still protect polarity, agency, dates, quantities, conditionality, quotation ownership and incompleteness. If two reviewers disagree, isolate the recording span and record the unresolved reading rather than voting to hide it. Make a short client-facing rationale describing how the choice balances accuracy, accessibility and readability.

A useful mastery test is to hand only the translated result to a new reader and ask what happened, who committed to what, and what is still unknown. If the answer differs materially from what careful listeners can establish in the source, revise. If the differences are only the predictable differences between a verbatim transcript and a disclosed cleaned edition, document that distinction. This is transfer: the method survives changes of genre because the semantic invariants are identified before surface editing.

Transfer laboratory 13: The conflicting editorial briefs

A client calls for ‘verbatim’ and ‘fully polished’ simultaneously. Show a sample paragraph where those objectives conflict, then propose an explicit two-layer delivery: a traceable source-aligned transcript plus a labelled readable version. Document the exceptions requiring full fidelity: corrections of numbers, identity, conditions, permission and contradiction. Record the concrete edit decision, the phrase that justifies it and the audience that will receive the result. Then carry out a three-way review with the source audio, an unedited source transcript and the translated deliverable. Do not let a successful style review substitute for evidence that the final commitment, sequence and qualification survive.

For this laboratory, prepare a second output with the opposite editorial purpose, then compare what legitimately changes and what must remain invariant. The target may compress ordinary planning pauses for readable communication, yet it must still protect polarity, agency, dates, quantities, conditionality, quotation ownership and incompleteness. If two reviewers disagree, isolate the recording span and record the unresolved reading rather than voting to hide it. Make a short client-facing rationale describing how the choice balances accuracy, accessibility and readability.

A useful mastery test is to hand only the translated result to a new reader and ask what happened, who committed to what, and what is still unknown. If the answer differs materially from what careful listeners can establish in the source, revise. If the differences are only the predictable differences between a verbatim transcript and a disclosed cleaned edition, document that distinction. This is transfer: the method survives changes of genre because the semantic invariants are identified before surface editing.

Transfer laboratory 14: The interrupted consent exchange

Someone begins ‘I can agree if—’ and another speaker interrupts. Write the transcript with the unfinished condition. The summary must not state that consent was given, and the translated dialogue must not guess the missing contingency. If the decision is consequential, flag the recording as insufficient rather than smoothing the statement. Record the concrete edit decision, the phrase that justifies it and the audience that will receive the result. Then carry out a three-way review with the source audio, an unedited source transcript and the translated deliverable. Do not let a successful style review substitute for evidence that the final commitment, sequence and qualification survive.

For this laboratory, prepare a second output with the opposite editorial purpose, then compare what legitimately changes and what must remain invariant. The target may compress ordinary planning pauses for readable communication, yet it must still protect polarity, agency, dates, quantities, conditionality, quotation ownership and incompleteness. If two reviewers disagree, isolate the recording span and record the unresolved reading rather than voting to hide it. Make a short client-facing rationale describing how the choice balances accuracy, accessibility and readability.

A useful mastery test is to hand only the translated result to a new reader and ask what happened, who committed to what, and what is still unknown. If the answer differs materially from what careful listeners can establish in the source, revise. If the differences are only the predictable differences between a verbatim transcript and a disclosed cleaned edition, document that distinction. This is transfer: the method survives changes of genre because the semantic invariants are identified before surface editing.

Transfer laboratory 15: The data and privacy request

A speaker starts ‘the records were erased’ and corrects to ‘flagged for deletion’. Create a public notice that accurately describes process status. Explain why removing the repair does not allow the editor to retain the first action, and why even the second wording should not imply completed erasure. Record the concrete edit decision, the phrase that justifies it and the audience that will receive the result. Then carry out a three-way review with the source audio, an unedited source transcript and the translated deliverable. Do not let a successful style review substitute for evidence that the final commitment, sequence and qualification survive.

For this laboratory, prepare a second output with the opposite editorial purpose, then compare what legitimately changes and what must remain invariant. The target may compress ordinary planning pauses for readable communication, yet it must still protect polarity, agency, dates, quantities, conditionality, quotation ownership and incompleteness. If two reviewers disagree, isolate the recording span and record the unresolved reading rather than voting to hide it. Make a short client-facing rationale describing how the choice balances accuracy, accessibility and readability.

A useful mastery test is to hand only the translated result to a new reader and ask what happened, who committed to what, and what is still unknown. If the answer differs materially from what careful listeners can establish in the source, revise. If the differences are only the predictable differences between a verbatim transcript and a disclosed cleaned edition, document that distinction. This is transfer: the method survives changes of genre because the semantic invariants are identified before surface editing.

Transfer laboratory 16: The long-form question and its restart

An interviewer begins asking whether an action happened, then restarts to ask when a related review occurred. Maintain question type and the lack of an answer to the original formulation. Produce a brief multilingual QA prompt asking bilingual reviewers to identify which question the respondent actually heard and answered. Record the concrete edit decision, the phrase that justifies it and the audience that will receive the result. Then carry out a three-way review with the source audio, an unedited source transcript and the translated deliverable. Do not let a successful style review substitute for evidence that the final commitment, sequence and qualification survive.

For this laboratory, prepare a second output with the opposite editorial purpose, then compare what legitimately changes and what must remain invariant. The target may compress ordinary planning pauses for readable communication, yet it must still protect polarity, agency, dates, quantities, conditionality, quotation ownership and incompleteness. If two reviewers disagree, isolate the recording span and record the unresolved reading rather than voting to hide it. Make a short client-facing rationale describing how the choice balances accuracy, accessibility and readability.

A useful mastery test is to hand only the translated result to a new reader and ask what happened, who committed to what, and what is still unknown. If the answer differs materially from what careful listeners can establish in the source, revise. If the differences are only the predictable differences between a verbatim transcript and a disclosed cleaned edition, document that distinction. This is transfer: the method survives changes of genre because the semantic invariants are identified before surface editing.

Twenty-four-weekday exercises condensed into a four-week practice plan

Week one: detecting the event

Day 1 identify fillers and pauses; Day 2 distinguish lexical hedges; Day 3 mark abandoned numbers; Day 4 examine negation repairs; Day 5 practise speaker turn boundaries; Day 6 review raw ASR errors; Day 7 independently check ten audio segments. Keep source, markup, target and error note in one sheet. Maintain a cumulative audit log of repaired propositions and accepted target-language alternatives. At the end of each practice day, use the reverse-claim test: read only the target translation and list the assertions it implies; then compare the list to the final commitments the speaker actually made. If a mistaken assertion survives, identify the first transformation stage at which it appeared, and redesign that stage rather than merely tweaking the final wording.

This practice is suitable for translation students, bilingual reviewers, caption editors, localization staff and people evaluating speech-to-text systems. The exercises are educational; they do not confer certification, and formal legal, medical or specialist records require their own governing rules. Measure progress through fewer unmarked repairs, cleaner uncertainty labels and better independent reviewer agreement, not through speed alone.

Week two: protecting commitments

Day 8 compare corrected agents; Day 9 corrected objects; Day 10 temporal anchors; Day 11 conditionality; Day 12 quotation ownership; Day 13 uncertainty and probability; Day 14 reconcile a second listener’s independent notes. The pass condition is no unmarked change to who did what, when, under what conditions. Maintain a cumulative audit log of repaired propositions and accepted target-language alternatives. At the end of each practice day, use the reverse-claim test: read only the target translation and list the assertions it implies; then compare the list to the final commitments the speaker actually made. If a mistaken assertion survives, identify the first transformation stage at which it appeared, and redesign that stage rather than merely tweaking the final wording.

This practice is suitable for translation students, bilingual reviewers, caption editors, localization staff and people evaluating speech-to-text systems. The exercises are educational; they do not confer certification, and formal legal, medical or specialist records require their own governing rules. Measure progress through fewer unmarked repairs, cleaner uncertainty labels and better independent reviewer agreement, not through speed alone.

Week three: adapting outputs

Day 15 build a verbatim convention; Day 16 make a cleaned reading edition; Day 17 create timed subtitles; Day 18 audit accessible sound cues; Day 19 interpret a simulated repair; Day 20 draft a labelled summary; Day 21 compare all formats against the same recording and explain every deletion. Maintain a cumulative audit log of repaired propositions and accepted target-language alternatives. At the end of each practice day, use the reverse-claim test: read only the target translation and list the assertions it implies; then compare the list to the final commitments the speaker actually made. If a mistaken assertion survives, identify the first transformation stage at which it appeared, and redesign that stage rather than merely tweaking the final wording.

This practice is suitable for translation students, bilingual reviewers, caption editors, localization staff and people evaluating speech-to-text systems. The exercises are educational; they do not confer certification, and formal legal, medical or specialist records require their own governing rules. Measure progress through fewer unmarked repairs, cleaner uncertainty labels and better independent reviewer agreement, not through speed alone.

Week four: adversarial quality assurance

Day 22 collect difficult negative and number cases; Day 23 challenge ASR punctuation; Day 24 assess interrupted claims; Day 25 review accent fairness; Day 26 perform a specialist terminology review; Day 27 conduct an independent bilingual audit; Day 28 release a sample package with a short conventions note, uncertainty log and corrected translation. Maintain a cumulative audit log of repaired propositions and accepted target-language alternatives. At the end of each practice day, use the reverse-claim test: read only the target translation and list the assertions it implies; then compare the list to the final commitments the speaker actually made. If a mistaken assertion survives, identify the first transformation stage at which it appeared, and redesign that stage rather than merely tweaking the final wording.

This practice is suitable for translation students, bilingual reviewers, caption editors, localization staff and people evaluating speech-to-text systems. The exercises are educational; they do not confer certification, and formal legal, medical or specialist records require their own governing rules. Measure progress through fewer unmarked repairs, cleaner uncertainty labels and better independent reviewer agreement, not through speed alone.

Glossary: the terms worth knowing

Disfluency

An observable feature of speech production such as a filled pause, repetition, restart or self-repair. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Filled pause

A sound such as ‘uh’ or ‘um’ occurring during speech planning or interaction. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Silent pause

A time interval without speech; interpretation depends on placement and context. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

False start

An initial phrase or word fragment abandoned before its projected completion. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Reparandum

The stretch of speech replaced or withdrawn during self-repair. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Editing interval

Material such as a hesitation or ‘sorry’ between abandoned and replacement stretches. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Repair

The speaker’s replacement or correction following a trouble source. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Reparans

The replacement wording supplied after a reparandum. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Retracing

Repeating or returning to an earlier part of an utterance. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Turn-holding

A practice that helps a speaker maintain the conversational floor during planning. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Overlap

Two or more speakers talking at the same time. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Cut-off

A word or utterance terminated before completion. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Aphesis versus cut-off

Do not confuse linguistic reduction within words with an interrupted word fragment. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Self-correction

A speaker changes their own prior or incipient wording. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Other-initiated repair

An interaction partner signals difficulty and prompts correction or clarification; distinct from an intra-turn restart. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Fluency normalization

Editing speech to remove nonessential hesitation or repetition for readability. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Clean verbatim

A deliverable convention that commonly preserves substantive speech while omitting specified routine disfluencies. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Full verbatim

A deliverable convention retaining a broader record of audible spoken details; exact notation depends on project rules. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Diarization

Identification of which speaker produced which part of audio. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Source-aligned target

A translated representation traceable to the corresponding source span. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Epistemic modality

Linguistic marking of possibility, probability or evidential commitment. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Deontic modality

Linguistic marking of permission, obligation or necessity. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Polarity

The affirmative or negative status of a proposition. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Proposition

The content that can be asserted, denied, questioned or otherwise committed to. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Commitment

What a speaker asserts, requests, refuses or endorses at the conclusion of the relevant speech act. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

ASR

Automatic speech recognition, which converts acoustic signals to text and requires checking. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Subtitle cue

A time-aligned on-screen textual unit in a subtitled video. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Caption

Synchronized text representing dialogue and relevant sound information for accessibility. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Bilingual audit

Source–target comparison by a reviewer competent to evaluate the relevant languages. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Editorial provenance

A record of source version, editing decisions, notation, corrections and review. The label does not by itself determine a universal translation rule; its role is established by the source event, audience and agreed publication convention.

Frequently asked questions about translating disfluency

FAQ 1: Should I translate every ‘um’ and ‘uh’?

Not automatically. In a detailed transcript, the filler may be relevant evidence of pacing, floor management or a sequence under study. In a readable interview translation, many ordinary fillers can be reduced if the editorial method is clear and the final commitment is preserved. Do not turn a filler into an unsupported assertion of uncertainty or nervousness. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 2: Is a corrected number a factual change or a speech event?

It is first a speech event: a speaker produced one number and replaced it with another. Unless external evidence supports a real-world status change, do not describe it as an official amendment. In evidence-sensitive formats record the correction; in concise summaries use the final value with whatever verification is warranted. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 3: What is the difference between a false start and a complete retracted claim?

A false start may be a fragment never completed; a complete retracted claim may have been asserted and then withdrawn. This distinction affects how the audience interprets the speaker’s responsibility. Mark scope and completion before translating; an early fragment does not license the editor to invent its continuation. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 4: Can subtitles simplify repetition without violating accessibility?

They can often simplify nonessential repetitions for readability, but captions must convey the dialogue in essence and important sound context. W3C guidance acknowledges appropriate simplification while warning against omission of essential dialogue. Preserve consequential corrections, denials, speaker switches and sound cues, and keep timing aligned to the recording. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 5: How should I punctuate a self-correction?

Use the editorial conventions agreed for the project. A dash can show cut-off or repair; ellipses may show a trailing-off thought; bracketed notes may describe overlap or inaudibility. Do not use any mark inconsistently or let punctuation imply a claim was completed when it was not. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 6: How do I translate a sentence that never finishes?

Preserve its unfinished status. If the source remains open to more than one completion, a grammatical target version that selects one is invented content. The appropriate output might use a cut-off mark or an editorial note. A summary should say the speaker began a point but did not finish it, rather than supply the missing conclusion. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 7: What if a machine transcript is very fluent?

Fluency does not verify source accuracy. Check source audio for missing negation, reversed numbers, misheard speaker identity, omitted replacements and unsupported punctuation. Automated transcripts can be useful working drafts but should not be the sole basis for evidence-sensitive translation, particularly where a brief disfluency changes an instruction. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 8: Is hesitation evidence of deception?

A hesitation alone cannot substantiate a claim of deception. Speech planning, word retrieval, timing, turn management, discomfort and many other factors can produce pauses. A translator should represent the observable event and avoid inserting speculative motive. Interpretation belongs in a separately identified analysis where appropriate. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 9: Should I keep dialectal grammar?

Do not mistake a legitimate language variety for a speech defect or accidental restart. Consider the target audience, text type and speaker voice, and avoid invented nonstandard target forms that stigmatize or caricature the speaker. An oral-history project may preserve stylistic patterning; a neutral technical summary may emphasize factual content. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 10: Can I omit an abandoned word in a published quotation?

This depends on the publication’s quotation and editing conventions and whether removing it changes meaning. If the repair concerns a factual claim, number, agency or certainty, omission can mislead. For a clean paraphrase, do not present the result as an exact quotation; label it for what it is. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 11: What is the role of a second bilingual reviewer?

A fresh reviewer can identify omissions the first translator no longer notices, especially after reading a fluent edited version repeatedly. Provide audio, timecode, brief, glossary and unresolved issues. Ask reviewers to judge factual commitment and disfluency representation separately from style. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 12: How are interpreting and written translation different here?

Interpreting unfolds under latency and interactional pressure. The interpreter may render an early phrase before the speaker revises it; transparent target-language correction can be necessary. Written work generally allows replay and controlled editing. The same semantic invariants matter, but the repair mechanism is operationally different. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 13: What if the speaker switches language in the middle of a correction?

Record the switch at the point where it occurs and determine whether it introduces a term, names a person, quotes speech or repairs a misunderstanding. Choose established target terminology where evidence supports it, but do not infer a psychological or social motive from the switch alone. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 14: What if the first number is louder than the correction?

Prominence is not proof of final acceptance. Listen to the whole utterance and the interaction immediately following it. In the full transcript, annotate the audible first value and the correction; in a concise summary, report the corrected final value, preserving applicable uncertainty. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 15: How can an editor avoid making a speaker sound unintelligent?

Do not add pseudo-stutters, excessive filler spellings or exaggerated dialect markers. Avoid translating ordinary planning noises as inability, and use a consistent policy across speakers. Preserve distinctive voice when relevant, but separate factual accuracy, style and inferred competence. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 16: Does transcription standardization solve every translation choice?

No. ISO 24624 describes representation of recorded spoken interactions in a particular transcription framework, not a universal rule for every target language and editorial purpose. Translation decisions still require a brief, risk assessment, target-language expertise and review of what must remain recoverable. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 17: What does ISO 5060 contribute?

ISO 5060:2024 provides guidance for evaluating translation output, including analytic error-based evaluation and human evaluator competence. It is not a special interpreting standard and does not dictate which fillers every caption must show. A project can adapt analytic error categories to lost repairs and altered commitments. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 18: What is an evidence-sensitive transcript?

It is a record whose readers may need to check how something was said, not only what conclusion a summarizer drew. It needs a documented convention for pauses, replacements, uncertainty, speaker identity and source alignment proportionate to the project. Applicable formal rules should be obtained from relevant professionals. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 19: How should I handle a correction from ‘approved’ to ‘reviewed’?

Treat it as a change of action type. A clean target claiming approval is wrong if the speaker withdrew that verb. In a detailed transcript, show the repair; in a summary, say that the speaker reported review rather than approval. Do not infer approval from surrounding optimism. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 20: When is a sigh worth captioning?

When it is important to understanding the audio information or scene, it may warrant a cue. Avoid attaching a precise emotional interpretation such as ‘ashamed’ unless supported. A clean textual report may omit the sound while still preserving the substantive answer. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 21: Can I back-translate to test fidelity?

A reverse paraphrase can reveal whether a target states stronger certainty or different agency, but back-translation alone is not proof. Compare with the source recording and a competent bilingual judgment. Machine back-translation can reproduce the first model’s error. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 22: Why are timing and sequence relevant?

The listener hears a tentative or mistaken fragment before the repair, and the delay may affect turn-taking, suspense or evidence. Subtitles that reveal a correction too early can change the viewer’s inference. Research transcripts may need timecodes; summaries need not show every second. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 23: How do I flag ambiguous speech?

Separate ‘inaudible’ from ‘unfinished’ and ‘uncertain word’. Supply a source segment and confidence rationale only when evidence allows it. Never fill a missing clause merely to achieve grammatical completion or consistent punctuation. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 24: Who owns final editorial decisions?

The client and translation team should agree on deliverable purpose and conventions; qualified reviewers assess high-risk content, and specialist rules govern their domains. A translator may explain the tradeoff, but must not silently rewrite source commitments to suit a style preference. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 25: How can I practice efficiently?

Choose short authentic recordings you have permission to use, mark disfluency spans, draft both detailed and readable outputs, and ask a second listener to infer the commitments from each. Track wrong polarity, number, speaker, modality and incomplete-claim errors. Five careful passages provide more insight than hundreds of unchecked smooth sentences. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 26: Are AI speech translation systems improving at this?

Research continues on disfluency-aware correction and speech translation, including peer-reviewed work in 2026. The task has not been made risk-free: data, audio quality, language pair and output policy affect outcomes. Test the actual system on your own task-sensitive material instead of assuming fluent generations reflect every source repair. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 27: Does ‘verbatim’ mean phonetic imitation in another language?

A translation cannot duplicate every sound relation because target languages have different hesitation conventions and prosody. A verbatim-oriented translation aims for traceability of meaningful events and documented treatment of audible features. Mechanical copying can create a social effect that was not present in the source. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 28: What must survive when an utterance is summarized?

At minimum, the valid final proposition, its source ownership, degree of certainty, conditions, negation and any unresolved incompleteness. If a correction itself explains a dispute or establishes how information changed, it may need explicit mention. Summaries are compressions, not licences to fabricate clean certainty. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

FAQ 29: How does this chapter differ from conversation repair?

It treats within-turn production events such as fillers, abandoned words, false starts and self-corrections. The neighbouring conversation-repair chapter handles clarification, common ground and turn-by-turn negotiation of understanding. They intersect when a second speaker prompts repair, but their diagnostics and editorial priorities are distinct. The quality check is always evidence-based: examine what the speaker actually completed, whether a prior wording was withdrawn, and which information an intended reader must retain. A target that is grammatically smooth but changes the final commitment is not acceptable; a target that mechanically reproduces every hesitation while creating unnatural new implications may also fail. Record the editorial policy and let purpose determine a proportionate surface form.

Research references and careful limits of their use

• ISO 24624:2016, Language resource management — Transcription of spoken language — Published standard covering XML/TEI representation of recorded spoken interactions; reviewed and confirmed in 2022. It offers an approach to transcription, not a prescribed rewrite of all translations.
• ISO 5060:2024, Translation services — Evaluation of translation output — General guidance — Published standard on human evaluation of translation output, evaluator competence, analytic errors and sampling. Its scope excludes interpreting.
• W3C WAI, Captions/Subtitles — Accessibility guidance on synchronized media and the importance of conveying speech and meaningful audio information; automatic captions require checking.
• W3C WAI, Failure technique F8 — Notes that caption simplification can be appropriate, but omission of essential dialogue or sound cues undermines captions.
• Kumar, Gain and Ekbal, Mind the Pause (ACL 2026) — Recent peer-reviewed computational research on multilingual disfluency-aware speech correction; not an editorial mandate for evidence-sensitive translations.
• Bhat and colleagues, 2023 ACL Findings on disfluency correction — A research perspective on disfluency correction in speech and downstream translation; domain-specific methods require independent verification.
• Salesky and colleagues, Fluent Translations from Disfluent Speech (NAACL 2019) — A reminder that producing fluent translation from spontaneous input is a technical task distinct from keeping a complete record of what the source speaker uttered.

Research must be read within scope. A standard dealing with transcription markup is not a rule for speaker psychology; a guideline on translation evaluation does not prescribe interpretation conventions; a captioning accessibility resource does not require mechanically reproducing every hesitation; and a computational paper reporting improvements on test data does not guarantee acceptable results in legal, clinical or safety-critical speech. The examples here are constructed educational diagnostics. For real consequential work, consult the governing brief, accessible source recording, relevant field specialist and qualified language reviewer.

Next routes: where this skill fits in the complete translation system

For the full architecture, return to Master Art of Translation | The Complete System for Moving Meaning Between Languages. The predecessor, The Evaluative Stance Fidelity System, explains how praise, criticism, sentiment and intensity survive translation; this chapter adds the distinct problem of tracking a speaker’s abandoned or repaired wording. For inter-speaker clarification, visit The Conversation Repair Fidelity System; it addresses how participants resolve understanding, rather than making every within-turn hesitation its own interpersonal negotiation.

When the main difficulty is choosing accurate terms rather than tracing a spoken restart, consult the protected Vocabulary Learning Hub. When the reader wants a general explanation of meaning, English expression and their relation to other languages, use How English Works: Translation and Meaning Across Languages. These are outward references to established resources, not changes to their editorial ownership, titles or pages.

The final rule is simple enough to remember and demanding enough to practise: translate the accepted meaning, preserve consequential revisions, disclose permissible simplifications and never manufacture what the speaker did not complete. The best target does not pretend spontaneous speech was written in advance. It makes the audience understand what the speaker ultimately said, what was corrected and what remains genuinely unknown.