All five-stage completions kept headband below PSG for mean person macro F1 with valid outputs fixed in 100 BOAS people

We give an exact method to bound macro F1 when some predictions are missing. The reference labels and valid predictions stay fixed. Each missing prediction can take any class. We prove the method and test its endpoints by exhaustive enumeration. We then apply it to stored sleep labels from BOAS version 1.1.3. After pooling nights within each of 100 people, the mean headband-minus-PSG difference was between -6.131 and -0.560 percentage points. Every allowed five-stage completion gave a lower headband mean score in this fixed sample. A descriptive outer person-bootstrap interval was -8.022 to +0.707 points. Pooled epoch macro F1 and ordinary accuracy had completion bounds that crossed zero. Thus the fixed-sample result depends on the score and weights. It does not establish a ranking for new people. The package preserves the proof, code, original tables, endpoint labels, and audit records.

macro F1 · missing predictions · sleep staging · BOAS · PSG · headband EEG · exact bounds · person-level analysis