TRUST-ADAPT: Risk-Gated Reversible Test-Time Learning with Conformal Shift Evidence and Selective Memory

Test-time adaptation (TTA) can improve predictions after deployment, but the update itself can become a new failure mode when unlabeled streams are transient, contaminated, imbalanced, or recurrent. We introduce TRUST-ADAPT, a risk-gated TTA algorithm that treats adaptation as a reversible intervention rather than a default response. A frozen source model converts confidence nonconformity into split-conformal p-values; a batch-level Hoeffding gate triggers learning only when the fraction of anomalous samples exceeds a finite-sample threshold. Triggered batches are further filtered by predictive uncertainty, adapted with pseudo-label and entropy objectives regularized by an immutable source anchor and selective class-balanced memory, and accepted only if an anchor-risk validation check passes; otherwise the parameters are rolled back. Under super-uniform conformal p-values and conditional within-batch independence, the gate controls the probability of a false adaptation trigger at a user-specified level. We implement the complete procedure in PyTorch and execute a locked five-seed controlled benchmark on streaming digit classification with abrupt, recurring, and 15% contaminated shifts. Across 15 seed-scenario pairs, TRUST-ADAPT improves mean streaming accuracy over the frozen source by 3.20 percentage points (Holm-adjusted Wilcoxon p=0.0039) while using 49.8 update batches on average versus 90.0 for TENT-style adaptation. The method is particularly competitive under contamination, but recurring shifts remain a limitation. All reported values are generated by the accompanying executable code; no unexecuted numerical result is claimed.