Perfect Cuts

Gling-killer: transcript decides which take wins, waveform decides the exact frame to cut. One run drops a care package with every useful format — minimal decisions for the user.

Proven on real footage 2026-06-10 (v1→v5 iteration). The locked rules were learned from frames the user flagged — don't soften them.

The three locked rules (the product — never soften these)

1. Hard threshold in, soft threshold out. Clip START = first frame voice crosses -30dB (breaths and mouth noise live below it — a -38dB start grabs the inhale and reads as 2-4 dead frames; the user flagged exactly this). Clip END = where speech drops below -38dB (word tails are quiet; -30dB clips them). These exact values are frame-verified on reference footage and are ALWAYS tried first; the script falls back to per-clip calibration only if they produce a degenerate map (it says so when it happens).
2. Zero pad in, one frame out. in_frame = floor(onset × fps) — no safety pad; 0.12s of "safety" pad was flagged as "2-4 frames too long." out_frame = ceil(end × fps) + 1.
3. Intra-sentence silence ≥ 0.25s = suspect false start. Whisper MERGES restarts into one clean sentence (it deduplicates the repeated words), so the transcript alone hides them. The speech map splits blocks at ≥0.25s silences — when one "sentence" spans two blocks, assume the first block is an aborted attempt unless its text clearly continues into the next. Prefer the later attempt.

Workflow

0. Setup check (run silently before anything)