Get started
What is live
Each part of the model is scored once against a test plan that was locked before training. A part is Live only if it passed. Parts that failed are Preview (off by default, opt-in) or off. This table is generated from capabilities.json, which is generated from the results files.
| Capability | Status | Locked tests and numbers |
|---|---|---|
| Detection and tracking A box, a score and a track number for every animal, kept through the clip. | Pending | v1 (plan v2) T6: pending Detection always runs. Today it is the shipped Vitals cattle detector (and COCO YOLO11 for other species); bio-det-v1 replaces it only if T6 passes.
|
| Identity across days Which known animal a track is, matched against a gallery you keep. | Pending | v1 (plan v2) T1: pendingv0 (plan v1) T1: fail
|
| Pose 17 body points (AP-10K order) with a confidence each. | Live | v1 (plan v2) T2: pendingv0 (plan v1) T2: pass The v0 pose head is on: it passed plan v1 T2 and the file is byte-identical. The v1 head's test (plan v2 T2) is pending.
|
| Behaviour Lying, standing, walking, running, eating, drinking, grooming, per sampled frame. | Live | v1 (plan v2) T3: pendingv0 (plan v1) T3: pass The v0 behaviour head is on: it passed plan v1 T3 and the file is byte-identical. The v1 head's test (plan v2 T3) is pending.
|
| Forecast Where the animal's centre will be over the next 0.2 to 5 s, with a 1-sigma radius and an intent word. | Pending | v1 (plan v2) T4: pendingv0 (plan v1) T4: fail
|
| Held-out species Pose and forecast on animal families the model was not trained on. | Pending | v1 (plan v2) T5: pendingv0 (plan v1) T5: fail
|
| Lameness The lame state, from gait. | Pending | v1 (plan v2) T7: pending |
| Tiled detection Detection on 2x2 tiles for small, far animals (opt-in). | Preview / off | Cam plan v1 tiled flag bar: fail |
| Calm, stressed, in pain, likely unwell Welfare states from video. | Preview / off | no test exists |
PCK@0.1: share of body points within 10% of the animal's box size of the true point. Macro F1: the mean F1 over lying, standing and eating. Forecast errors are in box sizes; a negative paired difference means the model beat the baseline. Every interval is a 95% bootstrap interval. The full test plans and results are in the model card.