Code · 2026
Screening-room: a QC agent that watches the cut
An agent that watches and listens to a render end to end like a picky viewer and returns every defect with a timestamp and evidence.
10
audits per cut
- 10
- measured audits
- 5
- passes per screening
- 2
- instruments per number
Problem
Automated builds fail in ways humans notice instantly and scripts do not: a card that says “$ B” for its whole life, narration a quarter-second ahead of the picture, a chapter ending on black under speech.
Approach
Ten measured audits (text, framing, sync drift, dead air, loudness, level jumps, holds, hiss, deltas between frames, continuity) plus a transcript-against-picture pass, each producing timestamps, not opinions.
The craft
Verify with a different instrument: every number the agent reports is re-measured a second way before it is allowed to block a release, because the first version of the audits happily passed on empty inputs.
Result
Runs before every delivery of the finance and documentary channels; the blockers it caught are logged as the channel's taste standard.
View code
Stack
- Python
- ffmpeg
- Whisper
- Claude agents