Bounded demo
Try the physics-consistency check.
A bounded demo of the same VIDEO_CHECKS profile used on the held-out CogVideoX cohort. Pick a sample trajectory class to see what fires — no Cosmos, no upload required.
Lab · Physics-IQHeld-out writeupMethodology
Held-out result (protocol v1)
- Held-out flag rate
100%
held-out flag rate (n=9 · Wilson [0.701, 1.000])
- Failure mode
static_hover
all 9 clips · seeds 2–4
- Discovery separation
n=6
discovery kept separate (post-hoc)
Sample check
A bounded demo of the same VIDEO_CHECKS profile used on the held-out CogVideoX cohort. Pick a sample trajectory class to see what fires — no Cosmos, no upload required.
static_hoverTypical held-out CogVideoX clip: tracker stays airborne with near-zero frame-to-frame speed while the prompt requires a release and fall.
Fires static_hover on all 9 held-out clips (seeds 2–4 × left/center/right).
This surface scores canned trajectory classes that mirror the public reports. Full RGB → CoTracker3 scoring is not hosted here — published cohort numbers live on the Lab and in the held-out writeup.
Published evidence
Held-out flag rates, Wilson CIs, and detector thresholds are published on the Lab and in the technical article — no repo access required.
Lab · Physics-IQHeld-out writeupMethodology
Critique the detector or send a public artifact for a free bounded evaluation — see the article request for feedback.