AI Video Captions
Use the exact final narration as the caption text source and the approved voiceover as the timing source. Never replace complete captions with summaries.
Produce caption JSON, SRT, a narration-to-cue map and the HTML caption elements required by the composition. Show one cue at a time. Target 7–14 Chinese characters per cue and allow 16–18 only when a semantic unit should not be split. Use restrained keyword emphasis without changing the text.
Run scripts/verify-caption-coverage.ps1 against the canonical narration and final HTML. A mismatch blocks delivery. If narration or voice timing changes, regenerate timing and rerun the verifier.