phase2/phase2_report.md

Phase 2 — Motion-NDF window ablation report

⚠️ SYNTHETIC PROXY MODE — trained on the synthetic manifold (gated AMASS, docs/BLOCKERS.md). Demonstrates the mechanism (H3): why poses must be lifted to motion windows. Not the final scientific numbers — re-run on AMASS windows for those.

Window W = 9 (+ joint velocity).

AUROC(valid vs negative) — higher = better separation; 0.5 = blind

negative type	static W=1	Motion-NDF (W+vel)
shuffled (dynamics-wrong, same poses)	0.498	1.000
highvel (dynamics-wrong, fast)	1.000	1.000
pose_ood (off-manifold poses)	1.000	1.000

Takeaway

The static field sits near chance (0.5) on the dynamics-wrong negatives (shuffled / highvel) — the per-frame poses are identical, so a pose-only field cannot tell smooth motion from broken motion. Motion-NDF, seeing the window + velocity, separates them (gain shuffled +0.502, highvel +0.000). Both catch off-manifold poses. This is the justification for Phase 2 and motivates including velocity in the reward (Phase 1 use_velocity).