Results so far

What we have tested, what broke, and what held up. The short version: the grader catches broken data, never cries wolf on good data, and matches a method that costs a thousand times more. Full technical write-ups live with the code.

Also here. The Stera-10M and SABER-10K evaluation reports show the shape a delivered report takes on a customer corpus.

The five findings, in plain words

What we did	What happened
Graded 50 good datasets from the field's standard test collections	All 50 passed. Zero false alarms. When the grader says your data is broken, it is not guessing.
Deliberately broke datasets in measured steps, then graded them	The score fell in step with the damage, and the grader rejected the badly broken ones. It catches exactly the defect it claims to catch.
Checked a popular shortcut: can any score rank datasets by how skilled the human demonstrator was?	No. Not our one-minute check, and not the leading academic method after 17 hours of GPU time. Both agree on what can be measured, and both agree on what cannot. Anyone selling that ranking is selling a coincidence.
Hid a camera-controls timing bug in otherwise good data	Statistical checks cannot see it, ours included. So we built a dedicated 12-second detector that finds the exact size of the timing error, down to a single frame.
Graded a real community dataset end to end	It earned a clean pass: better than its scrambled version on 13 of 13 test clips, with prediction error cut by 74 percent. That is the report every submission gets.

The same story, as one picture

For the technically curious. Each panel is one of the findings above, with the actual measurements.

Four-panel results figure — **(a)** Scores of datasets made by skilled vs unskilled operators: the ordering flips from task to task, which is why no honest tool should rank "skill". **(b)** Damage vs score: only genuinely broken pairing between video and controls drags the score down, and heavy damage gets rejected outright. **(c)** Our one-minute check (blue) next to the 17-hour academic method (orange), reaching the same verdicts on the same data. **(d)** The timing-bug detector: every planted error found at its exact size.

Why "scrambled version" is the whole idea: training a big model on anything makes its numbers move, even on garbage. So Gantry trains once on your data and once on your data with the controls shuffled to the wrong clips, and only pays out if the real thing wins. Beating your own scramble is the one result that cannot be faked by training harder.