RECONSTRUCTED from the locked pipeline and configuration for the July 26 metered replication (claude-opus-5 via Claude Code 2.1.219, API billing). The original July 25 session prompt was not separately archived; original run was Claude Opus via the desktop app, exact version unrecorded.
# LOCKED PROMPT — The Cut Test (autonomous A-roll edit)
# Freeze this text for cost / re-run measurement. Do not improvise new rules.
# (Reconstructed from the locked pipeline + config used in the Jul 25 official run;
#  if you have the original chat paste, replace this file with that exact text.)

You are running an autonomous A-roll video edit. Cuts only. No color, no titles, no music.

## Inputs (already on disk — do not ask for the reference edit)

- Source media: `raw/The Cut Test Raw File One.mp4` (4-channel Mac recording is expected; speech may live on a subset of channels)
- Word-level transcript: `transcript/words.json` (WhisperX / umcheck format: `[{start,end,word}, ...]`)
- Editorial knobs (frozen): `config.json` / `tools/config.json`
- Pipeline implementation: `pipeline.py` / `tools/pipeline.py`

You do **not** receive:
- The human reference edit
- The reference decision log
- Post-hoc human corrections

## Goal

Produce a publishable-style talking-head cut that:
1. Removes dead air / silence beyond intentional short beats
2. Strips fillers (um/uh/…), leading discourse where configured, and bloopers
3. Drops false starts and abandoned takes (markers like “let me start over”)
4. Detects retakes of the same script beat and keeps the strongest survivor only
5. Never interleaves takes across multi-pass sections — choose one complete pass
6. Refines every cut boundary against the audio waveform
7. Renders a final MP4 with short fades and loudness normalization

## How to run

Prefer the existing pipeline rather than inventing a new one:

```bash
# From project root (~/cut-test-cost)
python3 tools/pipeline.py \
  --media "raw/The Cut Test Raw File One.mp4" \
  --transcript transcript/words.json \
  --outdir .
```

Or dry-run (cutlist + decisions only):

```bash
python3 tools/pipeline.py --dry-run \
  --media "raw/The Cut Test Raw File One.mp4" \
  --transcript transcript/words.json \
  --outdir .
```

Expected outputs:
- `cutlist/cuts.json` — kept segments with source in/out times and reasons
- `logs/decisions.md` — beat / retake / pass selection rationales
- `output/edit.mp4` — final render (when not dry-run)

Honor every value in `config.json` (silence gaps, padding, fillers, false-start markers, retake thresholds, render CRF/preset/loudnorm). Do not retune mid-run.

## Multi-channel audio

This source may be 4-channel. The pipeline’s silence gate and downmix logic already account for speech on a subset of channels. Do not discard the file or force mono upstream unless the pipeline requires it.

## Stop conditions

- One official autonomous pass only.
- When `output/edit.mp4` (or dry-run cutlist + log) is written successfully, stop.
- Do not open or compare against any “reference” or “timebolted” cut.
- Do not re-render for taste after the first successful final.

## Deliverables checklist

- [ ] `cutlist/cuts.json`
- [ ] `logs/decisions.md`
- [ ] `output/edit.mp4`
- [ ] Brief note of wall time, model, and any tool errors (no secret keys)