The review found one unsupported inference and several wording issues, so I revised `output/diagnosis.md`; the main conclusions are unchanged. The revised file is 487 words against the 500 limit, and the original is preserved unchanged.

## Change log

1. **Removed an unsupported inference.** The original said the audio peak warning would remain because 00:00:42 is outside the clip's 00:00:08–00:00:24 range. The screenshot only says `Used 00:00:08–00:00:24` and does not say whether that is a timeline position or source in/out points. It now says the help file does not link relinking to that warning, so treat it as still open.
2. **Step 3 no longer points at the recorded path.** The `Original location` path moved to "Still needs checking", labelled as the recorded path only, since the help file says the current location cannot be determined from a screenshot.
3. **Expected results are labelled "expected, not observed".** This covers the row no longer showing `Offline` and E204 no longer stopping a retry.
4. **Step 4 now follows the help file's wording** ("Select Burn into video") and the screenshot's `Available option` label. It also notes that what **Apply** applies is not shown.
5. **Step 2 notes the row already appears as `Selected: interview_camA.mp4`.**
6. **Findings reordered** so the secondary warning sits directly after the immediate blocker, with the caption setting marked as a separate, non-error change.
7. **Added one evidence line:** all three screenshots show `FrameCut 3.2`, the same version as the help file.

## Checks performed in this pass

- Re-read `inputs/operator-message.txt` and `inputs/product-help.md` and compared each claim against them and the three screenshots. I found no conflicts between sources: the Export panel's `Sidecar only` matches the Captions panel, and 00:01:00, 16:9 and MP4 match the operator's request.
- Ran a local script (`check.py`) that confirmed the word count, the six required sections, and that the screenshot labels my steps rely on appear verbatim in the revised file. This checks my text against my own transcription, not against the images programmatically.
- Confirmed the preserved original is the first submission (488 words, same as first counted).
- Confirmed the revised file makes no claim of having clicked, changed, relinked or exported anything.

## Remaining limitations

- This is screenshot interpretation only; no export was attempted or verified.
- The interview file's current location, the Relink dialog, and how **Burn into video** is selected are not shown.
- Whether the audio at 00:00:42 is distorted cannot be judged from a screenshot.

**Estimated human time to finish:** about 5–10 minutes to relink, switch the caption mode, export and review the file. This is an estimate, not a measurement.

Files are in `output/`:
- `diagnosis.md`
- `CHANGELOG.md`
- `original-submission/diagnosis.md`