BUD2 results (CRITERIA_BUD2.json, registered 2026-10-28; 216 runs, seeds 1220-1223)

Holdable 0.2 s pushes (y +-70, x +-50, xy +-50 N), 24 runs per arm. Zero-step holds / held / steps.

information in the hold-first entry and decision zero-step held steps
true force, 0 ms 22 24 3
true force, 10 ms 20 24 4
true force, 20 ms 8 22 24
true force, 30 ms 8 18 30
true force, 50 ms 8 20 34
true force, 100 ms 3 21 37
IMU-SME 11 23 15
tuned MOB, gain 30 8 18 34
tuned MOB, gain 120 0 21 36

Correction (2026-10-28, from the code, before BUDF): the delayed oracle replaced the estimate in the hold-first entry and the ankle-hold target (est[‘fast’]); the capture prediction that releases a step uses the spike-rejected copy of IMU-SME (est[‘fast_dec’] from dist.raw_med), which the oracle did not replace. The delay therefore acts on when and how the hold starts, not on the release prediction. The registration text (“entry and decision”) overstated this.