Watermelon In Easter Hay: the cymbal notation does not track the cymbal stem

A two-direction audio check of tab s35881, run against a circular-shift null, three positive controls and a five-stem negative control.

Built 2026-09-07 15:00 CDT · anchor clock: 97 snare anchors, piecewise linear, span 59.3s to 506.1s · tolerance 45 ms

The finding

1,778 notated cymbal events sit inside the scored span, and they match the cymbal stem at 1.14x chance forward and 1.27x chance backward. Both sit under the 1.5x floor this project uses for a real effect.
The same method scores snare notation against the snare stem at 5.74x and 5.28x. Kick reaches 3.00x, toms 3.67x. The method finds correspondence where correspondence exists.
The notated cymbal lane matches the SNARE stem better than its own. 2.14x against snare, 1.14x against cymbals. That is the shape of events written onto the backbeat grid rather than transcribed from audible cymbal strokes.
This is not a ceiling artifact. The dense notated set lifts the null, and there is still headroom to 6.0x forward and 3.1x backward. The lane reached 1.14x and 1.27x.

What it does NOT license

It does not justify leaving the ghost marks stripped. The 1,317 anti-accent flags are the original author's own notation. Stripping another contributor's marks was wrong on provenance grounds, and that stays true whether or not the lane tracks the audio.
It is a fact about the tab's transcription quality, not a licence to edit. No cymbal event was added, removed or moved on the strength of this measurement.
The upload still repairs real damage. The live tab currently shows all 9 track names as empty strings, and the verified file carries all 137 CDATA blocks.

Direction A and B across five detection thresholds

Measured Direction A asks whether every written cymbal event has audible support. Direction B asks whether every clearly audible cymbal onset has notation. Null is 200 circular shifts of the whole notated set inside the scored span.

flux pct	onsets found	density	A written→audible	A null	A ratio	B audible→written	B null	B ratio
80	2,478	5.55/s	0.476	0.444	1.07x	0.382	0.318	1.20x
90	1,533	3.43/s	0.311	0.282	1.10x	0.391	0.321	1.22x
95	929	2.08/s	0.193	0.174	1.11x	0.402	0.322	1.25x
97.5	576	1.29/s	0.132	0.109	1.21x	0.422	0.321	1.32x
99	307	0.69/s	0.070	0.062	1.13x	0.414	0.321	1.29x

Every cell in both ratio columns falls below 1.5x. The answer does not depend on where the detection threshold is set.

Positive controls: does the method work at all

Measured Same clock, same tolerance, same null, each notation lane against its own stem at the 95th percentile.

notation lane	stem	notated	onsets	A ratio	B ratio
snare (midi 38,40)	snare	193	466	5.90x	5.23x
kick (midi 35,36)	kick	107	652	3.00x	3.28x
toms (midi 41-50)	toms	6	569	3.67x	3.45x
cymbal (midi 49,51)	cymbals	1,778	929	1.16x	1.26x

Kick is the load-bearing control. Its anchors are not the snare anchors that built the clock, its band is 30 to 180 Hz rather than 2 to 12 kHz, and it still reaches 3.00x. The clock is sound and the method discriminates.

Negative control: each lane against stems it should not match

Measured If the cymbal lane scored no better against cymbals than against bass, it is at chance. It does not, and the row that stands out is not the one expected.

notation lane	vs stem	A ratio	B ratio
cymbal	cymbals	1.14x	1.27x	its own stem
cymbal	snare	2.14x	1.07x
cymbal	kick	1.12x	1.26x
cymbal	bass	0.87x	0.98x
cymbal	piano	1.03x	1.08x
snare	cymbals	2.32x	2.32x
snare	snare	5.74x	5.28x	its own stem
snare	kick	0.32x	0.37x
snare	bass	1.28x	1.31x
snare	piano	1.08x	1.16x

Snare notation behaves correctly: 5.74x on its own stem, 2.32x on cymbals through bleed, 0.32x against kick because snare and kick alternate. Cymbal notation does not: its best score, 2.14x, lands on the snare stem.

The submission artifact, settled from the files

Measured Chord elements and CDATA blocks counted in each candidate's score.gpif.

file	CDATA	Chord elements	verdict
PRESWEEP-r7715683.gp (saved original)	137	8	the reference
RESTORED-s35881-TEXTSAFE-v2.gp	137	8	exact parity with the original, ship this
RESTORED-s35881-TEXTSAFE.gp	136	7	one CDATA and one chord short
r8768414-LIVE-EXPORT.gp (the live tab)	136	6	two chords lost on the live tab
RESTORED-s35881-v3-with-chord.gp	0	7	quarantined, wipes all 9 names

The chord question, answered

The chord did not vanish through a new transcription decision. The saved original carries 8 chord elements and the live tab carries 6, so two were lost on the tab. TEXTSAFE-v2 already restores both, and its chord-definition ids and beat-level chord references show an empty symmetric difference against the saved original. The separate chord patch is redundant and poisoned: 7 chords rather than 8, and zero CDATA.

What the upload changes, against the live tab

positions added        0
positions removed      0
dynamic marks added    0
dynamic marks removed  0
dynamics changed       0
anti-accent added      1,317
note events    4,855 -> 4,855

The upload does exactly one thing. It restores the author's 1,317 ghost flags and repairs the 9 empty track names as a consequence of preserving CDATA. Nothing moves.

The one remaining step

File: RESTORED-s35881-TEXTSAFE-v2.gp
sha256: 87020f8c4243a7441db8187b4d7356b300e62b4d013787343de5ea4a0cfe27fd
Live tab read this session: artist Frank Zappa, latest revision 8908034, all 9 track names empty strings.
The press. Songsterr exposes no server-side submit, and publishing over another contributor's tab is Brandon's act. Create a copy to edit, rename it Watermelon In Easter Hay Brandon edit, import the file above, submit there.

Method

onset function   spectral flux, 1024-pt Hann, 256 hop, 44.1 kHz
                 cymbals/snare 2-12 kHz, kick 30-180 Hz, toms 80 Hz-3 kHz
peak picking     local maxima above the Nth flux percentile, 60 ms min separation
clock            piecewise linear through 97 accepted snare anchors (np.interp)
                 scoring restricted to the anchor span, 59.33 s to 506.08 s
tolerance        45 ms half-window
null             200 circular shifts of the entire notated set inside the span
tools            tools/cymbal_two_way.py, tools/lane_control.py, tools/cross_stem.py
data             data/cymbal_two_way.json, data/lane_control.json, data/cross_stem.json

What this did not test

Whether Demucs separation leaks snare energy into the cymbal stem strongly enough to explain the 2.14x snare result. The bleed direction was checked and snare-vs-cymbals scores 2.32x, so bleed is present and real.
Individual cymbal events. Every figure here is a rate over the whole span, and no single stroke was adjudicated.
Hi-hat lanes, held out of the cymbal class deliberately.