Answers are grounded in this paper and cite their sources.
Develops a theoretical framework for black-box acceptance decisions in auditing, suggesting conditions for admissibility.
This paper develops a mathematical theory of black-box acceptance under metadata-conditioned auditing, where the deciding institution cannot inspect internal model states and must rely only on publicly auditable evidence such as challenge responses, replay outcomes, certificates, provenance claims, logs, and bounded statistics. The framework is institution-relative: admissibility is defined with respect to declared good and bad classes together with declared false-accept and false-reject bounds. The central structural object is a common-law control state, which packages pre-audit metadata and exogenous randomization while having the same law under both hypotheses of a fixed local comparison pair and remaining independent of the pairwise signal experiment. Within this framework, the paper proves two exact necessity theorems. The first lower-bounds the selector-averaged sum of witness-level Kullback–Leibler divergences under exact audited witness factorization. The second gives an analogous lower bound on provenance-class source information under exact classwise source factorization and a common-kernel map to the audited observation. The paper also formalizes certified support conditions, selector legality, non-vacuity boundaries, pair-uniform allowance screens, count thresholds, budget blockage conditions, and approximate necessity theorems with auditable slack budgets. The results do not provide an optimal audit design, a world-truth theorem, or a full adaptive sensing theory. They instead provide a rigorous screening calculus for determining when metadata-conditioned black-box audits can or cannot support admissible acceptance decisions under declared audit-law assumptions.
No takes yet. Share an insight, caveat, or question.
K Takahashi (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: