A 5% union error limit per night allowed either macro recall ranking at 10% review in BOAS v1.1.3

Sleep-label review gains depended on the allowed pattern of reviewer errors. We tested fixed BOAS version 1.1.3 labels from 128 nights and 100 people. Both rules used a 10% expected review budget per night. Perfect review gave stage benefit scores a 1.64 point advantage over error risk scores in macro recall gain. A common reviewer model allowed arbitrary errors within a 5% limit in each night's review union. The sharp contrast bounds were -1.14 to +4.99 percentage points. A stronger 5% error limit at every epoch kept the advantage positive, from +1.39 to +1.98 points. An added check with a 5% error limit under each rule permitted -0.67 to +4.34 points. These bounds apply to fixed reference labels and specified models. They do not measure reviewer error or clinical effects. The package preserves code, source data, a proof, bounds, error allocations, 2,000 refitted person bootstrap samples, and reproduction commands. All 773 source files matched fresh downloads. A full repeat matched all 20 result files.

sleep staging · review sensitivity · macro recall · reviewer error · sharp bounds · artificial intelligence