Known BOAS labels forced opposite headband and human N3 orders in 447 of 4,950 person pairs

We tested person order from stored sleep labels in 128 records from 100 BOAS people. Known labels stayed fixed. Each missing human or model label could receive any of five valid stages. The score was mean N3 minutes per record, with equal record weights within each person. Exact endpoint checks certified opposite headband and human orders for 447 of 4,950 person pairs, or 9.03%. Every allowed completion retained these opposite orders. The pairs involved 79 people. For 151 pairs, both opposite score differences stayed at least 10 minutes. A common strictly increasing map of person mean device scores preserves these orders. Each pair certificate is exact. Their sum is a lower bound, not an established sharp global minimum. The result concerns fixed archived labels, not physiological truth or a new population. Separate code matched all 54,450 pair checks across 11 settings. A fresh download verified all 773 source files. The replay matched six numerical and verification files byte for byte. The package preserves source bytes, hashes, code, proofs, pair tables, figures, and repeat instructions.

sleep staging · N3 · headband · polysomnography · person-level order · partial identification · missing labels · BOAS · OpenNeuro