timscheuerai/content-vault/shorts-cut

shorts-cut

Stage 1 of shorts-edit. The judgment-heavy stage.

The rule

Each recording is ONE short, recorded with many attempts per beat.

For each beat keep the LAST attempt — but only one that actually finishes.

Aborted fragments (trailing off into "So…", "and…", a cut-off word like "succes-") are not attempts. Skip them and fall back to the last attempt that completes.

This matters. Taking "last" literally ships broken sentences. In the reference session one clip had seven attempts at a beat and the last three all failed. A blind "last attempt" rule would have shipped a garbled step 5.

Retakes often happen inside one phrase, separated by a sub-0.5s breath — so cuts must be made at word boundaries from the word-level view, not at phrase boundaries.

How

python $S/words_dump.py        # word-level view with [GAP] markers
# -> spawn one sub-agent per clip, in parallel, with the rule above
# -> each writes edit/edl_parts/<clip>.json
python $S/tighten_edl.py       # pass 2 -> edit/edl_parts_tight/  (SOURCE OF TRUTH)

EDL shape: {"source": "<name>", "ranges": [{"start": s, "end": e, "beat": "...", "reason": "..."}]} Cut mechanics: start = first kept word − 0.06s; end = last kept word + 0.15s.

Pass 2 (`tighten_edl.py`) removes dead air inside a kept take — the speaker pausing mid-thought. That was 19.1s across 14 clips. It splits each range at acoustic silence (≥0.6s at −60dB), leaving 120ms of air. Cutting inside silence can never clip a word, so this needs no transcript at all.

THE TRAP: ASR word END times are fiction

This is the single most expensive lesson in the whole pipeline.

Scribe pads a word's end deep into the following pause — one word claimed to run 1.3 seconds past where the voice actually stopped. It also emits phantom words over dead-silent audio.

Consequences, all measured:

• A word-gap scan reported 2.5s of internal silence where the audio had 19.1s.
• Checking cut edges against the word grid produced 25 "cuts inside a word" alarms.

Measuring the audio showed 24 were sitting in silence. Only one was real.

Word STARTs are reliable. Everything about silence comes from the audio.

Use word timings only to avoid cutting mid-word — never to find a pause.

The other trap: ASR merges repeated words

The speaker said "X", dragged for 0.8s, then said "X" again. ASR merged both into a single "XX" token — so the take-picker literally could not see the duplicate; it looked like one word.

Two follow-on effects:

1. The EDL kept a 0.5s segment that was just the first X.
2. Once that was dropped, the surviving X had no caption, because the merged token's

start time now sat outside the segment.

Fix: edit/word_fixes.json supports _insert — put a word back at the time it is actually spoken (in SOURCE seconds), and it flows through the EDL like any other word.

{"<clip>": {"_insert": [{"src_at": 256.68, "word": "X", "dur": 0.30}]}}

The same file corrects mis-hearings for CAPTIONS only (audio is never touched):

{"<clip>": {"mode": "moat"}}

Expect this much waste

55–75% of every raw take is false starts and dead air. If your cut is only removing 20%, you are not looking hard enough at the retakes.

Verify

python $S/verify_cuts.py — no cut edge may land on live speech. It requires both tests to fire: inside a word and voiced. See shorts-qa for why either alone is useless.