dfreq-roller-native-calibration-20261103

Native CPU robust search learned one shared command-feature bias of -0.056 m/s, confirmed over256 development starts. All inherited neural weights remain unchanged. This is a general memoryless calibration, not neural-network weight training.

Legal unchanged61 float32 observations ->14 float32 action offsets. Roller accessories,50Hz. No seed,clock,worldposition,per-start lookup or privileged policy inputs are added. Model SHA256: 891db5713885fba1009ae926c9010c010ef1f9c4771bd3990d84a639469444d8.

Local unchanged original Arena evaluation at command1.1: 2.727s,5/5 finishes; separate128/128 held-out finishes and64/64 stress finishes across commands1.0/1.2. These are local results, not a server acceptance or rank. Noninferiority gates passed; superiority was not established. The existing published2.570s specialist is faster on its prescribed starts; no new personal-best claim is made.

This is a substantive trained/optimized controller with verified behavioral differences from our previous submissions, not a metadata-only variation. The frozen policy and validation preceded publication. Original physics, preparation, actuators, observations and termination were retained.

Inherited policy lineage and rights are documented in upstream-base-card.md and LICENSE-NOTE.txt. Own updates do not relicense inherited weights. Simulation-only; no real-robot deployment or safety claim. Private training data, traces, checkpoints and credentials are excluded.