# Yunke — transcript timing provenance

Part of the [AI Voice Agent](/docs/modules_handbook/shared/voice-agent/readMe.md).

## What it does

Explains the native transcript offsets, their limitations and the management-page
warning. The transcript remains the vendor's original text and spans. Nothing here
rewrites a call, trims text, downloads audio, reanalyzes a result or places another call.

The revised PDF asks whether text can return through the API; it does not specify
millisecond audio alignment. The approved App/Web route has real returned-text and
same-call ingestion evidence. Estimated end offsets are a disclosed quality limitation,
not another API capability requirement or a reason to repeat a phone test solely for
alignment. Actual speech accuracy remains part of approved recording/pilot review.

## How it works

### Native estimate, not measured word boundaries

Existing APK 2.0.20 static evidence traced `DialogueReporter.b` through smali because
the Java decompiler produced a control-flow stub. Each native start is the chat object's
local timestamp minus the connection start. That chat timestamp is assigned when the
event is processed; it is not an independently measured audio-word boundary.

The reporter calls `mj3.a(text, "zh", 1.0)`. For this branch the estimator counts
non-whitespace UTF-16 units at 200 units/minute, or about 300 ms/unit. It adds that
duration to the start, then uses the next start if `1 < next_start < estimated_end`.
It does not cap the last end at the call duration. This applies to the native uploaded
dialogue, not a petaV3 calculation. Language/speed behavior in other versions must not
be inferred from this sample.

Offline reproduction against the already captured 21:58 two-way sample matched **all
19 original ends**. Its last agent segment has 134 non-whitespace UTF-16 units:

```text
144178 ms + 134 × 300 ms = 184378 ms
184378 ms - 156723 ms call duration = 27655 ms overrun
```

This explains the stored end. It does not prove all of that final text was played or
align it to recorded audio. Agent text can describe an utterance interrupted by hangup.
Do not manufacture a corrected end or claim a full utterance was heard from its text.

### Preserve results and show the qualification

`YunkeCallTranscript` retains original spans, turn text and start offsets, and records
`end_exceeds_duration`. Both result mappers already use this grammar. The result remains
eligible for normal explicit-ID processing; the flag does not turn a connected call
into a failed attempt, suppress genuine customer speech or cause a redial.

`YunkeCallPresentation` emits fixed safe warning copy only for a consistently bound,
recorded result with that known flag. Unknown raw issue values are not returned. Missing
or conflicting bindings retain their existing review state without a transcript warning.
`YunkeCallState` shows the qualification in compact rows and expanded details:

> Transcript timing needs review. Some text may not have been spoken in full.

Rendering adds no vendor request, queue dispatch, storage write or audio access. Original
start/end provenance remains in the recorded metadata. Precise alignment would require
separately approved audio review; none was performed for this offline work.

## Reference usage

Open an already recorded Yunke call with an overrun in AI Calls or either profile call
tab. The timing warning accompanies the recorded-result state even when collection and
after-call actions are paused. Opening the row does not retry, settle or fetch media.

The existing synthetic presentation regression checks safe copy, unchanged local rows
and no warning for a conflicting binding. Component checks cover compact/expanded
display without controls. The prior private research contains the APK method evidence
and 19-row formula reproduction; credentials, real IDs and transcript content stay out
of this handbook.

## Related files

- [YunkeCallTranscript](/src/VoiceAgent/Support/YunkeCallTranscript.php) — unchanged original span mapping.
- [YunkeCallPresentation](/src/VoiceAgent/Support/YunkeCallPresentation.php) and [YunkeCallState](/resources/js/Components/YunkeCallState.vue) — local warning projection/rendering.
- [Presentation tests](/tests/Feature/VoiceAgent/YunkeCallPresentationTest.php) and [component tests](/resources/js/Components/YunkeCallState.test.js).
- [Web results](/docs/modules_handbook/shared/voice-agent/yunke-web-results.md), [call views](/docs/modules_handbook/shared/voice-agent/yunke-call-views.md) and [owner acceptance](/docs/modules_handbook/shared/voice-agent/yunke-owner-acceptance.md).
