Pillar 2 · World-model verification
ShippedPhysics checker
A physics-violation detector you can audit — calibrated, not claimed.
Position-only physics-consistency checks — object permanence, ballistic consistency, contact plausibility — the same interface as tracked video from a generative world model. Validated the only honest way: against ground-truth trajectories with known, labeled violations. Calibration completes in ~5 seconds on a laptop.
What we tested
Can a position-only checker catch labeled physics violations without false positives under tracking noise?
- recall on 399 violated trajectories
1.000
recall on 399 violated trajectories
- false positives on 200 negatives
0
false positives on 200 negatives
- hover detected at 100% (3 frames)
60 ms
hover detected at 100% (3 frames)
Charts
Flag rate by violation class
flag rate
Calibration results
| Class | n | Flag rate | Detected by |
|---|---|---|---|
| Clean (must not flag) | 100 | 0.000 | — |
| Clean + 2 mm tracking noise | 100 | 0.000 | — |
| Teleportation (0.3–0.8 m) | 100 | 1.000 | permanence |
| Anti-gravity hover (10–20 frames) | 99 | 1.000 | ballistic |
| Causeless impulse (1.5–3 m/s) | 100 | 1.000 | ballistic |
| Interpenetration (3–10 cm) | 100 | 1.000 | contact |
Methodology
- Generate labeled trajectories. Synthesize clean and violated position trajectories (teleport, hover, impulse, penetration) with known ground truth.
- Noise control. Add a clean + 2 mm tracking-noise negative class to probe false positives under realistic jitter.
- Run VIDEO_CHECKS detectors. Score permanence, ballistic, and contact checks; record which detector fires per class.
- Report recall / FP. Aggregate flag rates; require perfect recall on violations and zero flags on negatives before any generative-model claim.
Seeds: 50 seeds across labeled violation and control classes
- negatives must not flag (clean, clean+noise)
- violations must flag at rate 1.000
Notes
Calibration proves the detector on known violations before any generative-model claim. Video path: real Physics-IQ quiet; CogVideoX multi-seed cohort flagged via static_hover (see Physics-IQ experiment). Synthetic proxies at 100% prove the pipeline can fire. Cosmos/NIM deferred where geo-blocked.