Trust-Gated, Model-Assisted Supervision of Deployment-Time Adaptive Filters: A Benchmark and an Empirical Account of Supervision Tiers

Adaptive filters run on billions of devices, and nearly all of them throttle their adaptation rate. Aggressive adaptation diverges when the plant drifts from its calibration, and in active noise control (ANC), our setting, one screech into a user's ear outweighs decibels of average-case gain. Learned policies could drive adaptation harder, but nothing bounds a misgeneralizing policy in a live signal path, and model-based safe-RL certificates need the plant knowledge that goes stale. We formalize safe deployment-time adaptation, contribute a GPU-vectorized headphone-ANC benchmark with ear re-seats and calibration faults, and measure what supervision costs under matched controls (same in-loop checkpoint, only the supervisor varies, plus an on-policy protocol trained natively per tier) against a hand-tuned watchdog, CUSUM and EWMA null-calibrated offline at equal false-alarm budget. Our supervisor, Evidence-Gated Adaptation (EGA), has two tiers. The trust-gated reflex tier (randomized anchor probes, certificate-free rollbacks, and a virtual-probe channel validated by an anytime-valid e-process) trails the on-policy watchdog by a small paired deficit on the main suite (-0.14 dB, CI [-0.19, -0.11]) but beats it under a polarity miscalibration fault (+0.70 dB, CI [+0.60, +0.81]). Under a duty-cycled harm adversary it ends 12.3% of episodes in severe harm, against 21.4/21.2% for budget-matched CUSUM/EWMA charts fed the same virtual evidence and 38.3% for the watchdog. The optional certified tier (mixture e-processes, trial-certified commits, demotion, and a logged harm ledger) costs a further +0.94 dB paired. It adds guarantees and an audit trail without lowering any harm rate. Benchmark and code are provided as anonymized supplementary material.