Physics-IQ

Same pipeline: real quiet, CogVideoX freeze/hover caught.

CoTracker3 + VIDEO_CHECKS score RGB folders the same way. Real Physics-IQ ball-and-block-fall: 0% flags. CogVideoX I2V discovery cohort (6 clips, center/left/right, seeds 0, 1): 100% — all clips fire `static_hover` (tracked object stays airborne but stops moving). CogVideoX I2V discovery cohort on Physics-IQ ball-and-block-fall; n=6 (center/left/right, seeds 0, 1); post-hoc `static_hover` — not held-out confirmation. Not Cosmos/NIM. VIDEO_CHECKS catch freeze/hover generative failures. Not a broad generative-quality verdict.

What we tested

Does frozen VIDEO_CHECKS stay quiet on real clips and confirm freeze/hover on a pre-registered CogVideoX held-out matrix?

real flag rate (1 clip)

100%

CogVideoX discovery flag rate (6 clips)

static_hover

documented generative failure mode

Charts

Flag rate by cohort

flag rate

Methodology

Freeze detector + matrix. Lock VIDEO_CHECKS / static_hover thresholds and the held-out seed×scenario matrix before generating confirmatory clips.
Track + score cohorts. Run CoTracker3 on real Physics-IQ, discovery, and held-out CogVideoX folders with the same frozen profile.
Publish separately. Report discovery and held-out flag rates with Wilson CIs; document failure mode (static_hover) without pooling cohorts.

Seeds: CogVideoX discovery: seeds 0, 1, center/left/right (n=6)

real clips expected quiet for this scene
held-out matrix scored only with frozen VIDEO_CHECKS
generative cohorts document static_hover when flagged

Public methodology

Notes

CogVideoX = THUDM/CogVideoX-5b-I2V, not Cosmos. discovery n=6 (center/left/right, seeds 0, 1); `static_hover` documents one generative failure mode — not a broad generative-quality verdict.

One freeze/hover failure mode on CogVideoX ball-and-block-fall — not Cosmos, not a broad generative-quality verdict. Discovery remains post-hoc; cite held-out for confirmatory claims.