What is running, and what is not

Trimbin is mid-build. This page is a status report rather than a product tour, because the alternative is showing something that looks finished and is not.

Running, and measured

• The measurement layer. Exposure, focus, stability, audio and freeze detection, computed with ffmpeg in the same pass that builds proxies. Tested against footage with faults planted at timecodes we chose — results here, and they are the only earned numbers on this site.
• The archive and its queries. ClickHouse with vector and text indexes, holding a generated corpus of several hundred thousand decisions. That corpus proves the queries stay fast; it is excluded from every accuracy figure, because a number computed over generated rows measures the generator.
• The deployment. Everything here is Terraform, deployed by push, with the whole environment reproducible from an empty project.

Running on real footage

Twelve takes from a published, openly licensed dataset have been through the whole pipeline: measured with ffmpeg, read for a slate, embedded, encoded, and compared by the panel. Open the workspace and look at any setup — every take is there with its reasons, including the ones that were not chosen.

• Upload, end to end. A clip uploaded to a signed URL is measured, encoded and playing through the CDN without anyone touching it.
• Slate reading. A board on the front of a clip becomes a scene, shot and take number. No board is said to be no board, and the take is left ungrouped rather than guessed at.
• The panel. Three specialists and a chief, on the footage. It found what measurement could not: a take that stops mid-sentence, a whip pan that breaks eye-trace, a wall crossing the foreground.

Every setup so far has come back below the review margin and been flagged for a person. That is the correct answer, not a failure to produce one: twelve competently shot takes have nothing technical separating them, and a system that manufactured a confident winner from that would be lying.

Not built, or not yet earned

• Any accuracy figure. The number stays null. It can only come from editors overriding the system, and none has yet — the dataset ships an empty annotation template, so there is no ground truth to score against either.
• The assembled cut, and EDL export
• Natural-language retrieval over the archive
• Upload from the browser for anyone but us
• Sign-in on this deployment, pending one console step for an OAuth client

Listing these is deliberate. A demo page that quietly omits what is missing is the same overclaim as an accuracy figure computed over invented data, and this project has already made that mistake once.

What the archive holds right now

Generated to demonstrate that queries stay fast over millions of rows. Excluded from every accuracy figure on this site.

Check it yourself

Nothing on this site is a rendering of numbers you cannot fetch directly.

• /api/public/eval — results against planted faults, the earned numbers
• /api/public/accuracy — editorial agreement, currently empty and saying so
• /api/public/scale — real and generated counted apart