Superseded on 2026-09-08

Four verdicts on this page were measured against a flat 56.000 bpm grid that drifts up to 6.29 seconds away from the recording. The alignment gate reading UNUSABLE at 0.336, the coda discrimination of 1.15x, the ride per-event figure of 1.46x and the tom pan figure of 0.73x are all properties of that grid.

A metronome-anchored clock built from the Moises click stem reaches a kick and snare match of 0.684 against a guitar control of 0.196, and it locates 81 drum attacks in bars 98 to 104 that this tab leaves silent. Read Why the Watermelon GP is not done before acting on anything below.

Watermelon In Easter Hay, reconciled and shipped

Four parallel sessions had four answers for one drum staff. This is the decision list, the measurement behind each one, and the single revision that went live.

r8970744 LIVE 63 defects removed s6857183 copy tab 2026-09-08

The eight decisions

Each row is settled. A later chat should read this rather than re-derive any of it.

The defect nobody had finished counting

A snare cannot be struck loudly and quietly at one instant. A kick cannot be struck twice at one instant either. The live staff carried sixty-three positions that asked for exactly that, each one a voice-1 note landing on the identical bar, position and drum as a voice-0 note.

The earlier pass found forty-nine of them and stopped at the snare. Sweeping every drum instead of one found twelve more on the kick and two on the high mid tom.

The provenance test

The question is whether the doubled note belongs to the original transcriber or arrived in our own later editing. Revision r7715683 answers it directly. That file holds 2,091 drum notes, every one in voice 0, across 193 distinct snare positions with zero stacked. At all sixty-three of the disputed positions it writes one plain stroke and nothing else.

The lineage that would have caused damage

Two finished builds were sitting on disk waiting to be uploaded. Measuring them against the live head is what stopped them.

v13 disagrees with itself about which drum is playing. That file stores drum identity twice, as a Midi property and as an InstrumentArticulation index into the track's own kit table, and the two disagree. Stated in both scopes, because GPIF deduplicates note definitions: 2 note ELEMENTS disagree, and those 2 are referenced from 173 note INSTANCES. Read one way its 191 toms are 45×4 and 48×187. Read the other way they are 41×62, 45×4, 47×111 and 48×14. The two readings reconcile exactly: 62 + 111 + 14 = 187, and the 62 and the 111 are the two defective elements. The live head and the shipped file each return zero disagreeing elements, so this is v13's defect and not a reading artifact.

Corrected twice on 2026-09-08. The first version reported the single-lane figure as if it were the whole story. It was one of the two stored fields. A peer session reading the articulation index got the other, and the file is what is inconsistent. The second gave 173 without saying 173 of what. The articulation is the field the notation renders, so its numbers are the ones to quote, and every count here now names its scope. On a percussion staff the Midi property carries a pitched value and the kit index lives in InstrumentArticulation, which tr_ledger.py:189 states in its own source.

The chart build never carried the kit rewrite at all, so uploading it would have taken the kick from 504 back to 107.

The ride, settled on the stem

202 ride notes present in the original are absent from the live tab. Brandon's reading was that the original transcription wrote a continuous ride through bars where Colaiuta actually paused. The cymbal stem agrees with him.

The null is the 88 bars that kept their ride, median band energy 49.2. Four of the removal bars fall inside the alignment anchors and can be tested.

The anchor span runs from stem 59.3 s to 506.1 s, so bars before it and the coda after it cannot be tested on this clock. They were left exactly as they stand.

Why the coda stays empty

The printed chart marks tom heads as PRESENT in bars 95 to 104. No count appears here and none should, since that scan is 5.0 px per staff space and a notehead count needs 15. Nothing was written into the tab there, and the reason is a measurement rather than a preference.

The bars the chart marks empty score the same as the bars it fills. A detector that cannot tell those apart cannot license adding a note, so the heads stay out until a rescan or a coda-local clock exists.

The resolution ceiling, applied to every chart reading here

One scan underlies every chart observation in this lineage, at 5.0 px per staff space. Per-head identity needs 6 and a notehead count needs 15, and magnifying to 8x or 16x adds nothing, because the information is not in the pixels.

Two observations were withdrawn on these grounds rather than defended. A stacked-notehead reading at chart bar 20 is gone entirely. The chart bar 86 check survives only as filled noteheads are present there, never as four of them and never on a named drum.

Gates the build had to clear

The set difference, not a count

Both sides read from Songsterr's CDN. The change is a pure subtraction: 63 positions present in r8968524 and absent from r8970744, and zero positions present in r8970744 and absent from r8968524. Every removed position sits in voice 1.

The full list is published beside this page as setdiff.json, keyed by bar, voice, position in quarters and drum. The bars it touches are 6, 7, 8, 10, 11, 12, 13, 16, 17, 22, 23, 24, 26, 27, 28, 30, 31, 32, 35, 36, 38, 39, 42, 44, 45, 47, 54, 55, 56, 58, 59, 60, 62, 63, 64, 78, 80, 84, 85, 90, 92 and 97.

The same gates, run against the real stems

The transcription-repair machinery had passed its synthesized fixtures without ever meeting a real stem. Here it is, run against the actual 15-stem Watermelon set and the exact shipped artifact sha256 bc5ce3de02153f39.

Sixteen instants still ask for three hands

Present in the live tab before this edit and after it, unchanged. Each is a tom written under an untouched ride ostinato with the snare on the same instant.

Resolving these means dropping one of the three lanes, and the ride lane cannot be adjudicated per event at 1.46x. Recorded rather than guessed.

The two bad note definitions, named

Counting the defect is weaker than naming it. Two note elements on track 8 of v13 claim one drum in their Midi property while their articulation renders another. Reproduced independently by two sessions from the same file.

The arithmetic closes both ways: 111 + 62 = 173, and 62 + 111 + 14 = 187, which is the whole hi-mid-tom count the Midi property reports. CHART-TOMS is the control at 14 elements and zero disagreements, both fields reading 43×46, 45×4, 48×100, 50×35 identically.

Four ways to read this file and get a clean wrong number

Between two sessions, every one of these ran without erroring and returned a confident figure.

Verified here: searching the Properties list for InstrumentArticulation finds it on 0 of 15 reachable notes, which is exactly the silent zero.

Per-head tom identity: both routes run, both closed

Naming which physical tom was struck stayed open all session. Two independent routes existed and neither had been run to a verdict. Both are now measured on the isolated tom stem, and both fail.

Source separation blends the toms in both dimensions. It preserves that a tom was struck and when, and it does not preserve which one.

The source layer, checked rather than assumed

• A higher-resolution master would not help. The stem source here is 44.1 kHz 16-bit with 2.56% of its energy above 20 kHz, so it is full-bandwidth and not lossy-derived. Drums live well below that ceiling, so 96 or 192 kHz adds nothing to separation.
• Commercially sold "stems" for this album are re-recorded by session musicians rather than the original multitrack, so they answer which tom a session player chose, never which one Colaiuta struck.
• The one artifact that would settle it is the original studio multitrack, with each tom on its own mic. It is not publicly released.

So per-head identity is not pending work. It is closed at a named layer, and the tab's tom lane assignments stay as the chart and the transcriber wrote them.

Counts reconciled under one scope

Most of the apparent disagreement between sessions is scope rather than substance. Every figure below is a Guitar Pro note INSTANCE on track 8, never a raw stem detection.

The 342 against 293 gap is entirely the 49 stacked positions. Counting instances and counting positions answers two different questions, and naming which one is in play removes the disagreement.

Proof