The Watermelon cymbal lane was a broken clock, and the metronome stem proves it

Watermelon In Easter Hay rebuilt on the Moises metronome stem, 2026-09-08. Detector unchanged from the cymbal pass: 1024-point Hann, 256 hop, band-limited positive spectral flux, 95th-percentile local peaks, 30 ms refractory, 25 ms one-to-one tolerance. Only the clock changed.

5 of 5 lanes separate cymbal 1.45x to 3.00x clock tested on a held-out lane no note edited

The headline finding of the earlier pass is overturned. That pass reported the 1,573-note cymbal lane running at 1.46 times chance and separating from a wrong recording at only 1.45, the single failing lane across nine songs. On the metronome clock the same 1,573 notes, scored by the same detector against the same stem, reach 3.17 times chance and separate at 3.00. The notes were always there. The ruler was wrong.

The clock, and how it earned the job

Moises ships a metronome track alongside the stems. On Watermelon it holds 997 clicks at a 539.9 ms median spacing, which is an eighth note at the score’s own 56 bpm, so two clicks to the quarter. The only unknown was which click is score quarter zero, and that offset was fitted on the snare lane, landing at click 51.

Fitting on snare leaves kick free as a test lane, which is exactly how the shipped alignment was built, so the two clocks can be compared on ground neither of them saw:

clock	kick notes confirmed	coverage	scored span
shipped snare-anchored alignment	226 of 504	44.8%	59.3-506.1 s
metronome clock	286 of 504	56.7%	41.3-528.5 s

The metronome wins by 60 kicks on the held-out lane and covers 41 more seconds of the song, so it is not trading span for accuracy. That is what licensed the rebuild.

Every lane, rescored

lane	written	onsets	confirm	chance	ratio	wrong rec.	separation	verdict	separation before
kick	504	909	56.7%	8.6%	6.56x	0.94x	6.95x	SEPARATES	7.49x
snare	342	691	56.4%	6.9%	8.20x	0.77x	10.67x	SEPARATES	8.86x
hat	246	549	51.6%	5.2%	9.86x	0.87x	11.36x	SEPARATES	not scored
tom	168	930	60.1%	8.7%	6.94x	1.54x	4.51x	SEPARATES	5.67x
cymbal	1,573	1,323	40.2%	12.7%	3.17x	1.06x	3.00x	SEPARATES	1.45x

Wrong-recording control is Keep It Greasey through the same Moises separator, unchanged from the earlier pass. The hat lane carries a verdict for the first time; the old cymbal convention excluded MIDI 42, 44 and 46 entirely.

What moved and what did not

Cymbal is the whole story. Confirm went from 19.3 to 40.2 percent on 1,573 notes, and separation from 1.45 to 3.00. A lane called unsupported is supported.
Kick and snare improved. Kick confirm 44.8 to 56.7 percent, snare 47.7 to 56.4.
Tom went the other way, 70.2 down to 60.1 percent confirm, while still separating at 4.51. A better clock overall can still be worse in one place, and that is worth saying rather than hiding.
No note was added, moved or deleted. This pass changed a ruler and nothing else.

The standing rule this settles. Use the Moises metronome stem as the first clock source wherever one exists, because it comes from outside the events being scored and cannot be tuned against them. Fall back to the score’s own tempo automation where the score demonstrably follows the recording, and to fitted anchors only when neither explains the audio. Never optimise tempo from a mixed drum stem.

The 101-collision comparison, rerun on the new clock

The earlier page selected 101 cymbal strokes that collide with a notated tom and asked whether they behave differently from the rest of the lane. That comparison was built on the old clock, so it had to be rerun before anything could be said about it. Everything except the clock is identical: same detector, same 25 ms tolerance, same CONFIRMED / AMBIGUOUS / ABSENT classifier, same selection with bars 10, 13 and 52 excluded, and it still resolves to 101 slots holding 101 note instances.

group	confirmed	rate	vs chance
the 101 colliding strokes	37 of 101	36.6%	2.83×
every other cymbal note	594 of 1472	40.4%	3.12×

The original conclusion survives. The collision criterion still selects nothing. Both rates roughly doubled when the clock improved, from 19.8 and 19.3 percent to 36.6 and 40.4, and the gap between them stayed noise. It is -3.7 percentage points at z = -0.74, well inside the |z| ≥ 2 a real difference would need. The sign even flipped, from the 101 sitting marginally above the rest to marginally below, which is what a null looks like.

The original page asserted this from a bare half-point gap with no test attached. It happened to be right, and it is right for a stated reason now. Chance on this clock is 12.9 percent against the 13.3 the old pass used. The lane total here reads 631 of 1,573 rather than the 632 on the rescore table, a one-note difference from scoring across the cymbal-only span instead of the all-events span.