| layer | check | verdict | rule / detail |
|---|---|---|---|
| representation | independent_decode | PASS | every chunk decodes with a valid CRC, legal codes, no sequence gaps; evidence hashes match 160/160 chunks |
| representation | known_answer_roundtrip | PASS | |decoded - source| <= step/2 on every known sample; unknown samples decode as unknown worst 0.500 steps (hand_left/keypoints_3d) |
| representation | range_and_presence | PASS | no value outside its declared ruler (encoder raises); every required field explicitly present (reader rejects absence) |
| measurement | calibration | REVIEW | a calibration reference exists for every measuring source camera UNCALIBRATED; metric channels rely on estimated intrinsics |
| measurement | independent_reference | REVIEW | visual estimates evaluated against a paired measured reference no paired suit/tracker/fixture capture for this session: accuracy unqualified |
| measurement | synchronisation | PASS | all channels on one declared clock with declared uncertainty single camera; media timeline +/- half a source frame |
| coverage | hand_keypoints | REVIEW | at least one hand's keypoints known on >= 30% of tracked frames hand_left/keypoints_2d 24% known, longest gap 26 samples; hand_right/keypoints_2d 6% known, longest gap 39 samples |
| coverage | contact_decidability | REVIEW | contact is decidable (hand and object both visible) on >= 30% of frames decidable 0%; 49 tracker 'no contact' flags on undecidable frames recorded as UNKNOWN |
| learning_fitness | role_separation | PASS | observations carry role OBSERVED; intent is INTENDED/DECLARED; outcome is a separate linked record 4 annotations |
| learning_fitness | availability | PASS | labels available only after their interval are flagged post-hoc and excluded from policy inputs 2 post-hoc segment labels |
| learning_fitness | consecutive_examples | REVIEW | >= 10 consecutive known hand samples for next-state examples longest run 4 |
| machine_fitness | machine_profile | NOT_APPLICABLE | frames, tools, workspace and capabilities of a target machine record carries no ACCEPTED/ACTUAL role |
| authority | consent | PASS | contributor consent recorded by the capture application given=True at 2026-09-30T05:50:37.623411Z; contributor pseudonymous id present; blur_applied=True |
| authority | release | REVIEW | a signed release permits this declared use no signed release is attached to this record; required before external distribution |
| id | route | estimator / device | native Hz | causal | qualification |
|---|---|---|---|---|---|
| cam.video | camera | h264 1600x1200 @29.9946fps, device android-a3c8da38ecda722d | 29.9946 | yes | raw capture; referenced by sha256 in evidence/source.json |
| cam.segmenter | camera | hand-detector+open-vocabulary-segmenter | 10 | yes | UNQUALIFIED: no paired reference capture (suit, tracker or f |
| cam.hand_pose | camera | hand-pose-21kpt | 10 | yes | UNQUALIFIED: no paired reference capture (suit, tracker or f |
| cam.hand_mesh | camera | hand-mesh-recovery-mano | 10 | yes | UNQUALIFIED: no paired reference capture (suit, tracker or f |
| cam.metric_depth | camera | monocular-metric-depth | 1 | yes | UNQUALIFIED: no paired reference capture (suit, tracker or f |
| derive.kinematics | derived | backward-difference | 10 | yes | |
| vlm.segmenter | camera | video-language-segmenter | 0 | no | UNQUALIFIED interpretation; reads the whole clip, so its lab |
| app.intent | authored | capture application | 0 | no |