When Is a Learned Command Adapter Worth It? Closed-Loop Identification and Counterfactual Auditing of Frozen Locomotion Policies

This paper proposes an adapter necessity audit for deciding whether a learned adapter adds recoverable value to a frozen command-conditioned locomotion policy. It separates same-state counterfactual headroom, deployment gain over a cross-fitted fixed action, and state-allocation gain over a frequency-matched randomized policy. On Go2, the confirmatory study uses 20 independent clusters for each of three query distributions and 200 full learner refits. At the stated thresholds, direct-control queries return NO-GO, while VGCC and MPC return ABSTAIN. A deployment-representative H1 audit also returns NO-GO, whereas a learner-level synthetic control returns GO.