Answers are grounded in this paper and cite their sources.
Adaptive diagnostic supervisors select evidence, screen probes, stop, and report diagnostic sets or refusals. Passive calibration can target a different score law from deployed stopped transcripts. We formulate an auditable calibration interface using each transcript as the sampling unit, with standard split-conformal validity under exchangeability. Finite-summary conditions and a score-distribution discrepancy bound characterize alternatives to full replay. Controlled studies compare replay, history stratification, and propensity correction, and expose unchanged-policy environmental drift. Offline adaptive acquisition of prerecorded hydraulic sensor channels provides measured-data evidence, not physical diagnostic intervention; residual-score replay falls below target in block-split evaluation. Calibration-budget and milder-fault aircraft simulations report class coverage, set categories, refusals, and acquisition costs. Independent episode maxima support a separate reporting-only rollout audit. The results establish a calibration protocol and failure modes, not operational diagnostic validation or physical-safety certification.
No takes yet. Share an insight, caveat, or question.
Heydari et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: