How a run works

1. Align (only with --audio). The external recording is measured against the video, and a master file is written with that recording as its audio.
2. Transcribe and analyse. The audio is transcribed once and cached. Silences, voiced spans and mouth noises are measured from the waveform.
3. Decide. Three judged stages choose the cuts: the cold open and outro, the retakes, and the content. The judge votes on candidates that the code generated. Code turns each pick into a cut, so the judge never writes a timestamp. A mechanical stage then handles pauses, fillers and mouth noises.
4. Render and check. Each cut’s edges are placed against the waveform: the level step across each join and voiced audio beside each edge are scored, then every join gets one refinement pass within 40 ms (refine_joins, before anything downstream reads the plan’s output times). One ffmpeg render produces the cut and the mastered audio. The output is then measured against the plan: duration, loudness, the first word, join quality, dead air and colour.
5. Overlay and report. An optional second pass adds the overlays with the audio copied, so the checks still hold. report.md records every decision.

The stage-by-stage map is in docs/architecture.md.