WhisperWeb MP4 product walkthrough — source and review notes
Created 2026-09-07 for ZL-166 / PR #26.

Permission and provenance
This sample was created specifically for this repository under the user's authorization
 to supply a demo. The script, workflow cards, and three-tone music arrangement are
original task output, available for reuse with this demo. There is no customer footage,
recorded human voice, stock music, or third-party product recording. Narration was
synthesized locally with eSpeak NG (en-us, 145 words/minute); no voice cloning.
The cards illustrate the upload/review/export workflow, not a screen recording.

Actual transcription run — 2026-09-08
Input: product-walkthrough.mp4, 14 seconds.
SHA-256: 581e2434bf687f58e3597c06d797c99e3c63be648ad7f786cb754761cc084f36
Provider: Replicate openai/whisper
Prediction: e8bxmypgssrgc0d0fst9x3f6mw
Version: 8099696689d249cf8b122d833c36ac3f75505c666a395ca40ef26f68e7d3d16e
Completed UTC: 2026-09-08T00:39:41.827043348Z
Settings: English, translate=false, transcription=srt, no initial prompt or script.
provider-result.json preserves the actual response and normalized parser output.
The run used the application's buildReplicateTranscriptionRequest,
getReplicatePredictionTarget and parseWhisperOutput functions. It was a direct
provider call, not an application task, account/billing or upload-path acceptance run.
Reproduction: load project provider environment and run
pnpm exec tsx scripts/transcribe-mp4-demo.ts /tmp/demo-provider-result.json
This creates a new provider prediction and incurs normal provider charges.

Editorial handling — NOT human listening review
The first two exported cues preserve the model's text and timestamps.
The model returned "renewed transcript" in cue 2; the synthesis script says
"reviewed transcript". This discrepancy has NOT been corrected by guessing.
The provider returned no cue for the final masked passage. An editorial
[unclear under background music] marker was added at 10.500–13.430 seconds,
using the existing source timing, so omission is visible. This marker is not
provider output or a human determination of audibility.
TXT/DOCX/SRT are generated from transcript.json with buildExportFile.

Human listen-review: PENDING
No independent human listening review has been supplied. P2-2 remains open.
A human reviewer must play the linked MP4, check both recognized cues and their
boundaries, resolve or flag "renewed/reviewed", and document the masked passage.
Record reviewer name, review date, source hash, per-cue findings and corrections.
Do not mark this sample reviewed or P2-2 passed until that evidence exists.