Coverage gaps: positions at which at least one engine is silent
These positions lie outside the three-way comparison. They arise mainly because the engines segment the music at different rates — AnalysisGNN emits a label far more often than musWM or AugmentedNet, as Ratings sets out — and, in a few works, because musWM follows the repeat-expanded order of performance while the other two remain on the written bars.
A gap in coverage is not a disagreement, and silence is not an error. An engine with no label at a position is neither compared nor graded there; folding these rows into the identical/different split would distort the agreement rate.