Community · E2 · artifact verified

Find video moments with contact-sheet judgments

Zero-embedding video search as MCP tools and an agent skill: 60-frame contact sheets are screened by one parallel Jev request each, top windows verified on full frames, and any coding agent gets moments, timestamps and probabilities - no vectors anywhere.

01 · Role in the system

What Jev does here

ffmpeg lays a video out at one frame per second onto labelled contact sheets (60 tiles per minute); one Jev request per sheet, all in parallel, screens the whole video - every question rides in the same request since output tokens are free - and the best windows are re-checked on full frames before the agent sees them. Typed lenses add custom per-minute questions for near nothing. On the CC-licensed film Tears of Steel (12:14), five recorded Claude Code sessions found a juice-box moment at 2:28.5 (p 0.95, 19 requests, $0.0054), read a newspaper headline off a frame, searched by reference photo (8 moments, 6 verified at 0.86-0.97), and mapped the whole film with an agent-written lens. Screening costs about $0.02-0.03 per hour of video; failed tool calls are left in the demo recordings.

02 · Control boundary

Where Jev sits

Judgment replaces the vector index: what counts as close is decided by the question asked now, not at index time. Contact-sheet screening exploits input-only pricing to judge every second cheaply, deterministic code thresholds probabilities and picks windows, and an MCP server plus skill hand the few verified moments to any coding agent as full frames.

Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.

03 · Known limits

What this evidence does not prove

  • Demos are author-run on one 12-minute CC-licensed film; the repo's own docs note embeddings remain better for repeatedly querying a large library.
  • Every jevreel query pays Jev for the minutes it screens; there is no persistent index to amortize.
  • 2 stars, created 2026-10-15; speech mode depends on captions or local speech-to-text.
  • Primitives also use the SDK's Box and Image types beyond choice/noul/score.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.