Watermelon In Easter Hay: the two-hands audit

Superseded 2026-09-08: the cymbal stem was asked directly

The repair list on this page is withdrawn as a repair list and stands as an investigation queue. Its central move was to treat a tom and a cymbal sharing a sixteenth as a collision a drummer's hands could not play, then convert 21 measures and 101 cymbal strokes into removals. Hand availability cannot settle whether a cymbal sounded. The isolated cymbal stem can, and it was run event by event.

• The 101 are not a group. They confirm against the cymbal stem at 19.8 percent. The other 1,472 cymbal notes confirm at 19.3 percent. The collision criterion selected nothing.
• The lane is the defect. All 1,573 cymbal notes run at 1.46 times chance on their own stem, where snare scores 6.89, kick 5.21 and toms 7.71 through identical machinery.
• Bleed control passed, so this is a real measurement of cymbals. Notated toms carrying no notated cymbal confirm at 9.2 percent, below the 13.3 percent chance floor.
• The sub-sixteenth regrid is contraindicated. All 49 fractional tom positions sit a median 15.4 ms from a tom-stem onset and 81.8 ms from the sixteenth grid, in bars where the clock is independently accurate to about 16 ms. Snapping them to sixteenths moves them away from the audio.
• Display bar 25 is confirmed at 19 of 19 pairwise orderings, 129.2 Hz down to 80.7 Hz. Display bar 57 is contradicted at 0 of 10, with ten strokes near 86 to 91 Hz followed by one at 113.0 Hz written as the lowest drum in the kit.

The snare figures on this page need their scope stated. 193 counts plain non-ghost snare instances. 342 counts every instance in r8968524, being 193 plain plus 149 restored ghosts. On the chart side this disk holds four reader passes returning 83, 136, 137 and 154 heads from one image. Until a comparison names its revision, ghost rule, bar range and reader pass, the chart-to-tab gap carries no information about what the drummer played.

Full method, controls and per-event verdicts: the cymbal stem pass.

What the diff screenshot actually was

The preview page was Brandon's own editor change sitting on top of my upload r8968036. It published one minute later as r8968524, described "Fixed tom fills and rolls", at 17:43:08Z on 2026-09-08.

Diffing the two drum payloads measure by measure, exactly 2 measures differ: index 24 and 26. Songsterr's diff view labels systems by measure INDEX, so those print as 24 and 26 in the screenshot and display as bars 25 and 27 in the tab. The measure between them is byte-identical, matching the un-highlighted middle system.

What Brandon changed by hand

• Index 24, drum voice. A 32nd-note flurry with High Tom and Hi-Mid Tom stacked on one instant became plain sixteenths walking down Hi-Mid, Low-Mid, Low Floor.
• Index 24, cymbal voice. Twenty unbroken ride sixteenths became an eighth, a sixteenth rest, and a Low-Mid Tom written into that voice.
• Index 26, cymbal voice. Three snare strokes added at sixteenths 14, 15 and 16, two of them ghosts. The drum voice was left alone.

Why my upload had it wrong

1. Tom positions came from stem onsets and were never snapped to the chart's sixteenth

The printed Drumnet chart engraves at a sixteenth. My plan carried raw onset positions. In index 24 the ten planned onsets sat at 11.23, 11.55, 11.94, 12.85, 13.78, 14.74, 15.73, 16.01, 16.45 and 18.63 sixteenths. Two pairs are closer than a third of a sixteenth, so the notation writer subdivided to 32nds and stacked two toms on one instant.

Across the 40 tom-carrying measures, 30 hold a placement more than 0.20 of a sixteenth off grid and 14 hold two toms closer than three quarters of a sixteenth.

2. The cymbal voice was never re-marked against the fill

The chart pass wrote into the drum voice only, so the cymbal voice kept running. 21 measures of the live head have a cymbal voice with zero rests while three or more toms play underneath. The worst is index 56 with 11 toms and 11 cymbal strokes landing on the same sixteenths.

3. The missing snares are the chart reader's recorded blind spot

That read found 154 snare heads against the tab's 193, and the limit is already written into reference_watermelon_drumnet_chart_read.md. Busy measures lose heads.

4. Low-Mid Tom and Low Floor Tom were never available to the pipeline

The chart carries three tom lanes, mapping to High Tom, Hi-Mid Tom and High Floor Tom. A fill that walks down four drums cannot come out of that map.

The crop, and what it does not prove

Chart bar 16 is the measure that becomes tab index 24. This is a native-resolution crop of page 1, system 7, second bar, x 272 to 552 and y 722 to 788 of the 595x842 scan, magnified 8x with no downscale anywhere in the path. Ten filled oval noteheads follow the run of X heads, matching the ten events the plan wrote for this measure.

The instrument is NOT established here

A notehead shape plus a staff position plus the tab's own InstrumentSet is one witness, notation. The independent witness would be the isolated stem. It was tested and it fails.

Over the fill window 117.857 to 119.732 s, the toms stem sits at 16.94x the energy of the surrounding context. Against a null of 4000 random windows of the same 1.875 s width drawn from the same stem across the whole 545 s, the median ratio is 0.69, the 95th percentile is 26.45 and the maximum is 271. p = 0.061, so 6.1 percent of random windows match or beat it. That does not clear a 5 percent bar. The tom stem is mostly silence with sparse bursts, so a window that lands on any burst looks this strong.

Conclusion: the count of ten heads is established on the native crop. Naming those ten strokes as toms rests on notation alone at this bar. The +9 chart-to-tab offset remains confirmed at the SONG level, rank 1 of 96 rotations, p = 0.0104. A per-bar instrument attribution is a different claim and it is not carried by that result.

What the stems do show in that window

Every other drum stem drops inside the window, cymbals to 0.83, snare to 0.53, kick to 0.67. The cymbal figure is the one that bears on this audit: the cymbal is quieter through the fill, and my upload wrote twenty unbroken sixteenths across it. That ratio carries no null test of its own and is reported as an observation.

The repair list, 21 measures

Every row is measured from the live r8968524 payload. "Collisions" counts cymbal strokes sitting on a sixteenth a tom already occupies, which is where a hand has to leave the cymbal.

Sub-sixteenth tom beats, 41 of them

Position and length are in sixteenths inside the measure, zero-based from the downbeat.

What is not done, and the blocker

No corrected file was written and nothing was uploaded. Brandon is hand-editing the live tab right now, and r8968524 is his work. A machine pass over the whole part would overwrite it. The local upload_ready/watermelon-v9.gp is one revision behind the live head, so a corrected build needs a fresh export of r8968524 first.