Metrics

Two families live here and they do not agree with each other: frame_accuracy() scores the raw per-frame probability curve, while everything else scores a decoded event list. The disagreement is large, runs in both directions, and has four independent causes — docs/frame_vs_event_metrics.md is the measured account, and the reference for which numbers to quote.

f_measure	F-measure metrics: overall beat F-measure and per-position (downbeat) F-measure.
continuity	Continuity metrics: how much of a track is tracked at the right metrical level, versus at any self-consistent one.
confusion	Phase-confusion metric: half-cycle "1" <-> opposite-position swap rate.
position_accuracy	Bar-position accuracy, decomposed by how a track's predicted bar grid is misaligned with the reference rather than just how often it is.
frame_accuracy	Training-time metrics computed directly from per-frame probability curves.
tempo_acc1	MIREX Accuracy 1: octave-tolerant tempo accuracy.