Jevclip

Yes. Place same-named SRT, VTT, whisper JSON or timestamped text next to the video; videos without subtitles are listed and skipped rather than guessed. Plain manuscripts still yield judgments without highlight cuts.

Code applies value, chatter, hype and focus thresholds to Jev probabilities, and every drop carries a reason. Use export and eval on labeled segments to pick a threshold that fits your footage.

Mostly. The summary may only cite kept segment ids that code converts to times, unknown ids are removed and noted, every kept segment must be cited with a second rewrite pass, and numbers plus English names get checked.

No. Judgments are hashed by full content and cached in a local database, so reruns, threshold edits and budget changes reuse results. Only changed subtitles, focus points or wording trigger new calls.

macOS or Linux, Python 3.9 or newer, ffmpeg, and a TypeSafe API key. Install with uv tool install . from a clone, or fall back to python3 -m venv ~/.jevclip and pip install ., which puts the command at ~/.jevclip/bin/jevclip. Writing the summary also wants an OpenAI-compatible model endpoint, though without one the summary falls back to source excerpts. Upgrade with git pull and uv tool install --force ., remove with uv tool uninstall jevclip, and the judgment cache in ~/.jevclip/cache.db survives both.

One folder per video, jevclip-out/<video-name-short-hash>/, holding report.md with the timestamped summary, source excerpts and a per-segment timeline of why each was kept or dropped, plus highlights.mp4 capped at 180 seconds by default and built only from whole segments. full.mp4 is the de-watered cut and is not produced at all when no segment needs removing. Then highlights.srt and full.srt realigned to the cut videos, segments.json with each segment's type, probabilities, value score and reasons, and highlights.json and full.json mapping clips back to the original.

The synthetic 9m22s, 12-segment demo on jev-1.13.0 with default 20-second segments scored all 12 blocks correctly by second, with the hype block at 可疑说法 0.99, for 12 requests, about 17,000 input tokens and $0.0007. On a real 16m52s Chinese technical talk, the 45-second default had averaged the DOOM demo together with two filler segments to 0.49 and dropped it, while 20-second segments kept it and raised highlights from 2:24 to 2:59, at 32 requests and $0.0019 against 18 and $0.0012.