From Divergence Scores to Auditable Evidence: Exact-TV Anchors for Reward-Interface Dependence

Trustworthy paired-rollout audits require an observable behavioral target and a conclusion that survives independent administration. We meet both requirements for reward-aware agents by measuring how a frozen policy's action distribution responds to task-preserving edits of shaping meters, verifier messages, judge feedback, and reward-model outputs. The Paired Interface Response Audit (PIRA) admits edits through ground-truth preservation tests, estimates an intervention-indexed response profile from matched rollouts, and returns a finite-sample rank-conformal Holm decision (PIRA-Holm) together with a learned ranking score (PIRA-L). Three open-weight exact-TV anchors expose the target action-distribution shift directly; frozen reruns and low-overlap external banks then test execution and bank-design invariance. PIRA-Holm attains $0.872/0.868/0.866$ AUROC and PIRA-L attains $0.903\pm .012/0.897\pm .013/0.894\pm .013$ on $\tau$-bench-Exact, SWE-Gym-Exact, and MLE-Bench-Exact. Independent operators reproduce these results within $0.005/0.006/0.006$ AUROC. Three full external handoffs with $8.6$–$11.2$ semantic overlap reach $0.884\pm .015$, $0.882\pm .014$, and $0.879\pm .014$, retain worst-family AUROC of at least $0.802$, and keep Benign-Hard false-positive rates below $5$. These results turn reward-interface sensitivity from a score correlate into an exact-anchored interventional measurement and give paired-rollout audits a reproducible evidentiary bar.