Watermelon In Easter Hay, the Coda toms settled against the stems

Frank Zappa, Joe's Garage Act III. Songsterr s35881, drum track, Vinnie Colaiuta. Twenty-nine tom events were handed from the chart read to the stem session on 2026-09-08. This page gives every one of them a status. Built 2026-09-08.

events handed over 29 refuted 4 supported 2 firm, 4 weak no verdict 9 moot 10 kit stops at 538.9 s transfer gate REFUSED bar 25 clock bias -87.1 ms

RESOLUTION CEILING, and it limits every chart reading on this page. The Drumnet scan is 5.0 px per staff space. Per-head drum identity needs 6 px and a notehead COUNT needs 15 px. Magnifying to 8x adds no information. No notehead observation from these images is reportable as a finding, and that includes presence, so every notehead statement drawn from them is withdrawn.

What the scan DOES support is the non-notehead layer: rehearsal marks, printed bar numbers, time signatures, dynamics, passage text and the fermata. Those are symbols and words rather than heads on a lane, and they are legible at this scan. The withdrawal of the "refuted by silence" verdict for GP 103 and 104 rests on the fermata and the dynamic markings in the closing system, so it survives this ceiling intact.

The verdicts on this page rest on the stem measurements and their controls. The chart supplies the bar grid and the passage markings, and it is not carrying any head count here.

The question this page closes

The chart read produced 179 tom noteheads and handed 29 of them to the stem session for testing. That session returned verdicts on 20 of the 29 and marked the remaining 9 UNTESTABLE, because the alignment anchors stop at notation second 462.9 and the clock drifts by 165 to 773 milliseconds past that point.

Nine heads with no verdict is an open item, so this pass went after it from the stems. Four of the nine are now refuted, on a measurement that does not need a clock at all. The other five stay without a verdict, and the reason is now measured rather than assumed.

The headline, after the eye overturned it. The drum kit stops playing at 538.9 seconds, and that measurement stands. A first pass read that as refuting the four chart tom heads in GP bars 103 and 104, on a constant-tempo extrapolation that placed both bars after 539.7 s. Looking at the printed page withdrew that verdict. The chart's final system carries a fermata and a two-step drop to mp then pp, so the performance does not hold a constant tempo through the ending and those bar windows cannot be trusted. All four heads are printed as filled ovals. They return to no verdict.

Every one of the 29 events, with its status

Group	Heads	Status	Evidence
GP bar 25, the ten-note run	10	MOOT	Brandon wrote 8 of them by hand 26 minutes after the import. The stem supports 8 to 9 at every threshold. Nothing left to add.
GP bar 95	4	REFUTED	Position-matched test returns 0 of 4 at every threshold down to 0.2 percent of stem peak, where the detector finds 2,049 onsets across the song. Refuted for absence, not for quietness.
GP bar 97	6	4 WEAK	Four supported at 0.2 percent of peak, two of those also at 5 percent. A real find carrying weak evidence, and it should ship labelled that way.
GP bars 99, 100, 101	5	NO VERDICT	Bar-level tom presence does not discriminate these from chart-empty bars in the same region, at 1.15x. Clock drift exceeds one sixteenth, so no position test is available. See below.
GP bars 103, 104	4	NO VERDICT	All four confirmed by eye as filled ovals on the G line at 8x. A constant-tempo clock placed both bars after the kit stops at 538.9 s, and the chart's closing fermata invalidates that placement.

The one REFUSE was checked against the page as well. GP 95 is the only bar this project marks REFUSE, and that verdict came from a position-matched stem test returning 0 of 4 at every threshold down to 0.2 percent of peak. Looking at chart bar 86 at 8x, filled noteheads are plainly printed in that bar, so the transcriber did write them.

The two statements do not contradict each other and the eye governs the engraving. What the stem test can say is that no matching attack is present in the tom stem at those positions. What it cannot say is that the chart is empty there. REFUSE here means do not write these into the tab, and it does not mean the heads are absent from the page. The scan is 5.0 px per staff space, so the staff-geometry guard returns UNRESOLVED and nothing here names a drum.

chart bars 85 to 86 at 8x — The blue guide marks the lane the four GP 95 heads were read from. Filled noteheads are printed in the second bar. Passage text appears above the staff; which noteheads it governs is not established here.

Ten moot, four refuted, two firm, four weak, nine without a verdict. That accounts for all 29.

Where the kit stops, measured second by second

Peak band energy per second, taken straight off each stem with no threshold and no alignment. The kick falls from 48.3 to 0.016 across one second boundary. The toms fall from 17.5 to 0.41, and the hi-hat from 56.1 to 10.6.

stem second	toms	kick	snare	cymbals	hat	reading
535	20.43	34.33	0.005	16.02	0.10	kit playing
536	2.00	65.84	0.010	24.06	0.01	kit playing
537	17.54	48.29	0.000	11.49	25.34	kit playing
538	2.11	43.93	0.000	7.95	56.11	last bar of playing
539	0.41	0.016	0.000	3.70	10.56	kick and toms gone
540	0.14	0.016	0.000	3.14	1.82	cymbal ring-out only
541	0.021	0.016	0.000	7.51	1.27	cymbal ring-out only
543	0.00002	0.016	0.000	0.21	0.015	silence
545	0.005	0.024	0.005	0.04	0.010	end of file, 545.27 s

Where the clock put those bars, and why that is withdrawn. On the clock fitted to the last eight anchors, slope 1.01266, GP bar 103 spans stem 539.72 to 544.06 and GP bar 104 spans 544.06 to 545.27. Both would begin after the kick and toms have gone, and bar 105 extrapolates to 549.5 s against a file ending at 545.27 s. That slope assumes a constant tempo. The printed page shows a fermata over the closing bar plus a drop to mp then pp, so the ending is held and stretched by an unknown amount. The kit stopping at 538.9 s is measured and stands. Which notated bars sit under 539 to 545 s is not settled.

Why bars 99, 100 and 101 still have no verdict

The tempting move is to count tom attacks inside each bar and call a bar supported when the count is above zero. That test was built and it fails its own control.

The in-region control is the set of Coda bars the chart marks empty: GP 93, 94, 96, 98 and 102. Those bars should return nothing. They return as much as the claimed bars do.

threshold	chart-tom bars	chart-empty bars	ratio	result
5 percent of peak	3.00	2.60	1.15x	FAIL, under the 1.5x floor
2 percent	5.40	5.00	1.08x	FAIL
1 percent	9.00	8.20	1.10x	FAIL

Counting all seven claimed bars together, including the two dead ones, the ratio drops to 0.82x, which is below chance. There are real tom attacks scattered through the Coda, 17 of them at the strict 5 percent threshold between 506 and 541 seconds. They land in bars the chart marks empty as often as in bars it marks with toms.

A position-matched test would be the stronger instrument, and it is unavailable here. Clock drift in this region runs 165 to 773 milliseconds. One sixteenth note at the verified tempo is 267.9 milliseconds. The uncertainty is wider than the grid spacing, so a head cannot be assigned to its slot.

This is a closed answer rather than a gap. The five heads in bars 99, 100 and 101 cannot be confirmed or refuted from these stems, and the two measurements above are the reason. Settling them needs anchors measured inside the Coda itself, which needs a lane in that region carrying events both dense and unambiguous. The ride thins out and the snare stops at 511 seconds, so no such lane exists after bar 98.

The detector controls, and the two bugs found on the way

The final detector thresholds on a fraction of the song's global tom-flux peak, which is the protocol the stem session used. Window counts are then comparable to each other and to their table.

window	5%	2%	1%	0.5%	0.2%	role
541 to 545 s, song over	0	0	0	0	0	negative control, clean
496 to 501 s, toms quiet	0	0	1	1	2	negative control
486 to 496 s, toms audible	9	16	19	21	28	positive control
506 to 541 s, the Coda	17	33	61	96	131	the region under test

Two detectors were built wrong ahead of this one, and both are recorded

Median absolute deviation computed on a clipped residual. Negative values were set to zero, then the scale was taken, so more than half the samples were exactly zero, the deviation collapsed to zero, and the threshold became infinite. Every window returned zero onsets, including windows known to carry toms.
A staircase baseline minted its own peaks. The running median was sampled every 16 frames and held flat between samples, which is a step every 0.093 seconds. Subtracting a staircase leaves a sawtooth whose peaks sit on the step boundaries. The resulting Coda onsets arrived in clusters spaced 85 to 95 milliseconds apart, matching the stride and not the drummer.
A threshold cut from the population it filters. Scaling the threshold to local noise let a near-silent window out-detect every loud one, at 4.20 onsets per second against 3.40 in a window where toms are plainly audible. Replacing it with a fraction of the global peak fixed the ordering and made the controls behave.

One correction to the record

The notation clock for this tab runs at quarter equals 56. The tempo automation stores 56 against a half-note unit, and reading that as a quarter-note value doubles the tempo. Bar 93 is a 4/4 bar spanning 447.857 to 452.143 seconds in notation time, which is 1.0714 seconds per quarter and confirms 56 directly.

At 56 the 105 bars occupy 8 minutes 26 seconds, which fits a track running 9 minutes 5 seconds. At 112 they would occupy 4 minutes 13 seconds, which does not. One sixteenth note is 267.9 milliseconds, and any earlier figure of 133.9 is the doubled reading.

Three gates the first version of this page never ran

The first build tested 29 chart tom heads and stopped there. It did not put the repaired artifact through the transfer gate, it used a global tempo rule rather than local calibration, and it never matched cymbals one to one. All three ran on 2026-09-08 and all three returned something.

1. The transfer gate REFUSES the repaired artifact

Ledger built from the actual artifact diff, RESTORED-s35881-TEXTSAFE-v2.gp against CHART-TOMS-s35881-Brandon-edit.gp, scope set to drum track index 8. The diff is 179 added tom instances, matching the chart read exactly at 46 on midi 43, 98 on midi 48, 35 on midi 50.

Check	Result
artifact hash matches the file tested	PASS 9737d4ef629ca775
both artifacts are materialized files	PASS
no edit is still PROPOSED	PASS
every applied edit reached VALIDATED	FAIL 179 applied, 0 validated
at least one VALIDATED edit exists	FAIL 0 validated
alignment CONFIRMED ahead of any stem timing use	FAIL UNRESOLVED past GP bar 97
every voice passes the two-hands check	PASS max simultaneity 3, histogram 1653 / 307 / 1
regression: nothing outside scope moved, something inside did	PASS 2765 out-of-scope instances identical, 179 in scope

Verdict: REFUSED, no promotion available. Five checks pass and three fail. The playability check was run separately with drum_reality_gate.py and returned PASS, max simultaneity 3 voices across 2270 events. The three remaining failures reduce to one cause: 150 of the 179 heads have never been tested against the stems, so no edit can reach VALIDATED. Only 29 went through, being the bar 25 run and the 19 Coda heads.

Regression is clean. 2765 out-of-scope note instances are identical on both sides, meters and bar count and track names unchanged, and the only in-scope change is the 179 additions. Zero unrelated regression changes.

2. Bar 25 as local calibration, and it finds a clock bias

Brandon hand-wrote 8 tom notes in GP bar 25. Matched one to one against the tom stem under the mir_eval convention, so one transient cannot validate several written notes.

threshold	stem attacks	matched of 8	busiest window
5.0 percent of peak	15	7	1
2.0 percent	25	8	1
1.0 percent	32	8	1
0.5 percent	34	8	1

All eight signed residuals are negative: -119.6, -97.5, -58.0, -76.6, -112.6, -49.9, -56.9, -98.7 ms. Median -87.1 ms with a standard deviation of 25.3 ms. The median is more than twice the spread, so this is a constant clock offset rather than performance scatter. The stem attacks arrive 87 ms ahead of where the alignment map puts them.

What that settles. One sixteenth here is 267.9 ms. The raw median residual is 0.325 of a sixteenth, which looks like evidence for finer subdivisions. Removing the 87 ms bias leaves 25.3 ms of scatter, which is 0.094 of a sixteenth. The figure sits on the ordinary sixteenth grid. The competing sub-sixteenth tom counts of 41, 44, 49 and 51 are consistent with an uncorrected clock bias, and none of them supports minting thirty-seconds.

3. Cymbal one to one, and it licenses no removal

Fresh attacks read from the cymbal stem with tr_audio.cymbal_events, which separates a new stroke from a decaying resonant bed. Over the whole file: 1663 candidates, 908 NEW_ATTACK, 442 UNRESOLVED, against 1778 written cymbal events.

GP bar	written	fresh attacks	matched 1:1	reading
5	20	5	4	16 unmatched
14	15	8	7	8 unmatched
25	20	7	5	15 unmatched
85	20	9	6	14 unmatched
93	20	12	4	16 unmatched
95	20	13	9	11 unmatched
97	20	1	0	20 unmatched
99, 100, 101, 103	75	0	0	all unmatched

None of this authorizes deleting a cymbal. Global recall is 908 fresh attacks against 1778 written events, so the detector accounts for about half the notation across the whole song, and a per-bar shortfall reproduces that global rate. The skill permits removal only when a usable cymbal source affirmatively fails after alignment and leakage gates pass. The alignment gate is UNRESOLVED past GP bar 97, and 442 candidates are UNRESOLVED. Every unmatched written cymbal here stays UNRESOLVED and stays in the file.

The eye overrules the cymbal detector, and the chart says why. Looking at the printed page for these bars: chart bar 89 (GP 98) carries the Coda mark and (Gong), and chart bars 91 to 92 (GP 100 to 101) carry the passage text (cymbal on dome) at p above the staff. Both systems show a dense continuous run of noteheads on the top line, including a circled class left unnamed. Noteheads are plainly printed across all of those bars, and the tab writes 75 cymbal events there under its own articulation names.

The detector ran with instrument="ride" calibration. A dome stroke at p produces a much lower-novelty transient than a bowed ride, so a ride-tuned novelty floor classifies it DECAY_ONLY. The zero-fresh-attack result in GP bars 99 to 103 is a detector mis-calibration, and it is not evidence of absence. Deferring to the eye, every written cymbal in the Coda stands.

Correction on how those glyphs were named. An earlier draft of this page called the a notehead class after an instrument, on its shape alone. That names an instrument from a glyph, which this project forbids. What the page can carry is what the chart PRINTS in words: chart bar 89 is headed Coda with (Gong) beside it, and chart bars 91 to 92 are headed (cymbal on dome) at p. Those are the transcriber's own labels. The glyph shapes are recorded as a distinct notehead class and are left unnamed.

A second check, and it found my own measurement wrong. A circled notehead carries more ink than a plain x, so it could in principle cross the ink cut that separates toms from the ride on the top line. Re-measuring the raw page found no top-line head above 180 anywhere, which appeared to overturn the chart read's 35 top-line toms. The stored ink5 is inverted grayscale where high means dark, so the cut at 180 means raw 75 or darker. My re-measurement read raw grayscale, where a filled head sits at 35 to 69, so it could never exceed 180. The chart read is correct and my check was the faulty one.

The within-document control settles it. On the stored scale the 35 top-line heads run 185.6 to 219.9 with a median of 201.6. The kick control, a known filled notehead, runs 185.8 to 219.4 with a median of 200.8. The snare control is darker still at 219.8. The x noteheads sit at 104 to 127. The top-line heads carry the ink profile of a filled notehead, and their identity as a specific drum remains UNRESOLVED per the staff-geometry guard.

chart bars 89 to 90 at 8x — The top line runs continuous x noteheads across both bars. Two circled noteheads appear on the top line, logged as a distinct notehead class and left unnamed. The system also carries passage text above the staff. Which noteheads that text governs is not established here.

chart bars 91 to 92 at 8x — Dense top line with many circled noteheads, logged as a distinct notehead class and left unnamed. The system also carries passage text above the staff. Which noteheads that text governs is not established here. The detector found zero fresh attacks across these two bars.

The decision, which belongs to Brandon

Revision r8968524 stands on the copy tab s6857183. Nothing on this page has been uploaded anywhere. Three routes are open and the evidence now sits beside each one.

Route	What it means now
Leave r8968524 as it stands	The two firm heads in bar 97 stay out. The eight refuted heads stay out, which is correct. Nothing further is written.
Fold in what survived	Two firm heads in GP bar 97, plus four more there carrying weak evidence and labelled as weak. The eight refuted heads are dropped. The five without a verdict are held back.
Roll back to a toms-only file	Returns to CHART-TOMS-s35881-Brandon-edit.gp, which adds 179 chart toms and leaves voice 0 byte-identical. The kit rewrite in r8968524 goes away with it.

Two further items carry agreement from both sessions and neither is applied. GP bar 14 holds four tom heads that are x noteheads on the printed page, which the skip rule removes, and the stem supports only two of the six at 5 percent. GP bar 93 holds two heads that the stem supports at both 5 percent and 0.2 percent, and those look solid.