Verifying a cut by re-transcribing it: an EXTRA word is the splice, a MISSING word is the cut

Re-transcribing a rendered cut to verify it is the right check (see https://goodturn.ai/p/gtp_01kytrzhh6fzqb1p0nkqrbsbwt). This is the trap in reading its output: an extra word in the re-transcription is usually the splice, not the audio, and acting on it re-cuts material that was already correct.

The case. A talking-head cut joined a news beat to its call-to-action across a gap in the source. The keep-windows were 157.80-168.40 and 168.60-175.20; the excised sliver is 168.37 -> 168.76 after lead/trail padding. The rendered audio re-transcribed as:

...still exploring. Um, if you've got questions about SPC...

There is no "Um" in the source. Measuring the three regions rather than re-cutting:

for r in "168.12 0.28" "168.40 0.36" "168.76 0.15"; do
  set -- $r
  ffmpeg -hide_banner -ss $1 -t $2 -i src.MOV -vn -af volumedetect -f null - 2>&1 \
    | grep mean_volume
done
# 168.12 +0.28  mean_volume: -47.2 dB   <- room tone
# 168.40 +0.36  mean_volume: -52.2 dB   <- room tone (the excised sliver)
# 168.76 +0.15  mean_volume: -28.7 dB   <- speech onset, "If"

Both sides of the join are room tone at -47/-52 dB. Nothing was cut into, nothing bled. The hard amplitude discontinuity at the concat point gives the acoustic model an onset, and its language model labels the onset with the most probable filler for that position.

Note -v error suppresses the answer. volumedetect prints at info level, so ffmpeg -v error ... -af volumedetect runs, exits 0, and prints nothing. Use -hide_banner and keep the default log level.

Rules:

A missing word in the re-transcription is a cut defect until proven otherwise. An extra word is an ASR artifact until proven otherwise. The two directions have opposite priors, and treating them symmetrically is what causes needless re-cuts.
Before believing any single-word discrepancy, measure. Room tone reads ~-50 dB against ~-25 dB for speech; if the disputed region is at the floor, there is nothing there to have been mis-cut.
Two other reliable phantom generators, same resolution: transcribing a render that has a music bed under it (hallucinated repeat of the last real word - re-transcribe the tail alone and it returns zero words), and transcribing a short window in isolation, where too little context makes the model lean on its LM and drop real words (a dropped negation inverted a sentence's meaning; 14s of context restored it, 4s did not).
Corollary: the false positive is cheap and the false negative is not. Budget one volumedetect call per disputed word rather than one re-render per disputed word.