A 1% reference label edit cap per night allowed either stored model to lead in accuracy in 100 BOAS v1.1.3 people with fixed outputs

We derived exact bounds for paired accuracy when a shared categorical reference can change under an integer edit cap in each record. We tested the method with fixed PSG and headband model outputs in BOAS v1.1.3. The test used 128 nights from 100 people and 114,578 common valid epochs. Each person had equal weight. The original PSG lead was 0.418 percentage points. Its descriptive 95% person bootstrap interval was -0.257 to +1.085 points. At a 1% cap per night, unrestricted replacements gave bounds of -1.463 to +2.297 points. Replacements restricted to original adjacent human stages gave bounds of -1.343 to +2.074 points. Thus, both rules permitted either accuracy ranking. The exact first adjacent-stage reversal cap was 2/807, or 0.247831%, with 240 edits. That threshold applies to the primary weights. These are fixed-data sensitivity results. They do not show actual scorer errors or physiological truth. The package preserves the proof, reusable tool, source tables, fixed retrieval URLs and hashes, attaining edit sets, saved draws, and reproduction checks.

sleep staging · reference labels · accuracy comparison · sensitivity analysis · paired accuracy · BOAS · exact bounds