Is this grooming?

Some of these are controls, and the mix is not stated. If a clip is just a still or slowly-moving animal with no paw movement, that is a Not — and saying so is the only thing that makes the confirmation rate mean anything. A rater shown only candidates has no way to be wrong.

The detector selected windows where the body is still and head-region pixels move. Whether that is grooming, sniffing, chewing, a twitch or a tracking artefact is a claim about content, which no statistic in this project can settle. That is what this page is for.

Scored, and the answer was none of them. The panel was rated blind: 0 of 30 candidate clips are grooming, against a registered bar of 50%. None of the 15 speed-matched controls was called grooming either, so the panel was not scored indiscriminately. The detector does not detect grooming, and the spectral gate is not read as an answer about it.

What it fires on is tracking jitter. The skull triangle is rigid, so variation in its three bone lengths within a window is DLC noise and cannot be behaviour — and the detector never saw a bone length. Candidate windows carry 3.44× the skull jitter of their speed-matched partners (0.0313 against 0.0091 body lengths, non-overlapping, 26 animals). A still animal passes the first test, its keypoints jitter, the egocentric crop is aligned on those keypoints, and the crop shifting frame to frame reads as motion inside it.

The spectral gate had already refused separately — no arm of the registered grid reached its 500-candidate minimum — so these clips were never evidence for or against a 3–8 Hz rhythm. The panel stays open below: answers are kept in this browser and this site has no server, so nothing reaches anyone until you export.

Your scores

Send the exported file on, or drop it at work/grooming/confirmation.json in the pipeline repo — grooming_gate.py --phase combine reads it from there and publishes the confirmation rate as the detector's own precision.