dfreq-hill-nov1-general

Simulation-only Hill event controller:61float32 observations,14float32 radian-offset actions,50Hz, leg accessories. No new seed, clock, timetable, global position or observation-fingerprint lookup is used.

Native-physics derivative-free calibration learned a fast continuous controller. A second independently selected conservative blend uses delivered gravity and angular-rate feedback to apply at most3% of that controller's action difference from the incumbent. Both inherited neural backbones remain frozen. This is calibration/optimization, not new neural PPO training.

Frozen before fresh qualification: local unchanged native official sweep3.842s,5/5 finishes at command0.6, versus historical published3.889s. Fresh paired256starts:254candidate finishes versus253incumbent, no lost finishes, one rescue; shared-finisher mean gain0.039743s. Paired shared-finisher95%bootstrap interval [0.03432806324110671, 0.04513833992094861]. The two32start command stress sets (.55,.65) produced63candidate finishes versus62incumbent, no lost finishes. Finite sample evidence is not a universal safety guarantee, and no server result or points award is claimed.

Inherited weights and lineage are credited in upstream-base-card.md. Original neural weights are unchanged; new calibration does not relicense inherited material. Simulation only: inherited positive command clipping means zero command is not a validated stop. Do not use on a physical robot or as a general walker.

SHA256 4124ba296a4ce261847c4a6bf0cbca8e7f81abe7ed28c4bc3a212bd1d2018a32.