cut-auditlisted
Install: claude install-skill vaibhav-barman/live-clipper
# CUT AUDIT — the transcript is not evidence
> **THE ONE FACT:** an ASR language model is built to produce fluent text, so **two identical adjacent
> phrases decode as one**. A clean transcript proves nothing. In the original project, **five separate
> transcripts** called a cut clean that had two doubled lines in the first 20 seconds. A human found
> them by listening.
**Streamers restart lines constantly**, talk over game audio, and get interrupted by alerts and TTS
donations. Assume there are more problems than you found.
Adapted from the creator's skill. **Kept:** every gate and its measured threshold. **Changed:** all
energy reads run on the **Demucs voice stem** (`a.wav` from `shortform-cutaway`), never on the mix —
on the mix, a gunshot is a "syllable" and a music bed is "speech". Thresholds were measured on a
talking-head mic; **re-calibrate them on your first 3–5 stream clips** and note the new values here.
## The gates, in order
Run all of them. Each catches a class the others structurally cannot. Loop: any FAIL → back to the
cutter with the evidence, max 2 rounds, then flag the clip for a human.
### 1. Stretched-word screen (before beats are written)
Any word longer than ~1 s is an artifact: the aligner smeared one word across a hidden pause or a retake.
- Flag every word with `end − start > 1.0 s`; **at a clip EDGE, probe regardless of length**; for a
short function word, **> 0.55 s** is already suspicious.
- Resolve each against an **RMS energy map** (4