With fixed BOAS v1.1.3 labels and known masks, disagreement sampling reduced paired accuracy audit variance by 96.8% at a cap of 100 epochs per night

We tested an audit of the stored PSG minus headband accuracy gap in BOAS v1.1.3. The main frame contained 114,578 epochs from 128 nights and 100 people. Sampling only model disagreements reduced exact cohort audit variance by 96.78% at a cap of 100 reference labels per night. It selected 10,391 labels, versus 12,800 for uniform sampling. At cap 25, both designs selected 3,200 labels and the variance reduction was 90.20%. The weighted estimate was unbiased for the paired gap. A proof gives a universal variance bound for fixed scores that are zero outside a known set. Separate source parsing, exact subset checks, simulations, and a fresh download passed. The package contains source files, code, a prediction-only selection tool, results, hashes, and repeat commands. The result requires fixed labels, known eligibility masks, and independent random audit selections. It does not measure review time or uncertainty for a new population.

sleep-stage scoring · audit sampling · paired accuracy · model disagreement · finite population variance · BOAS