Exact Finite-Population Sequential Auditing for Selective Release from Model-Proposed Candidate Sets

Abstract

Predictive models can rank low-risk cases, but past accuracy cannot establish that skipping review is safe in the current batch. We formulate selective release as a risk-constrained finite-population audit design. A model-proposed candidate and its audit plan are frozen before candidate outcomes are revealed, after which uniform sampling without replacement leads to release, continued review, or full review. Exact hypergeometric recursion verifies the complete policy for every candidate defect total and evaluates its final verification demand. Across 675 generic settings, the maximum unsafe-release probability was 3.998% under a 5% limit. In a retrospective Xili-2026 development population (379,526 alarms; 504 defects), the final candidate covered 14.999% and contained two defects. At the primary 1% released-defect target, the target-specific plan required 16.55% fewer final candidate reviews than a one-look plan selected from the same pre-audit information. Candidate-level avoidance increased from 34.495% to 45.338%, while expected overall reviews avoided increased from 5.174% to 6.800%. These are operating characteristics under audit randomization, not observed factory savings; candidate construction and the conservative adjustment were selected during development.

The authors' abstract, as published at the source. Mathematics, 2026 · DOI ↗