Watermelon In Easter Hay. Songsterr copy s6857183, head revision r8968524. Event-by-event audio check of the 101 colliding cymbal strokes, run 2026-09-08 against the isolated cymbal stem. The lane-level conclusion on this page was superseded later the same day by a rebuild on the Moises metronome stem; see the banner below.
SUPERSEDED 2026-09-08, later the same day. Everything below about the cymbal lane being unsupported was measured on the wrong clock. Rebuilding the alignment on the Moises metronome stem, with this same detector, this same stem and these same 1,573 notes, moves the cymbal lane from 19.3 percent confirm to 40.2 percent, from 1.46 to 3.17 times chance, and from 1.45 to 3.00 times separation. The verdict flips from weak to SEPARATES, and all five lanes now clear. The notes were always there. The ruler was wrong.
What was wrong, precisely. This page anchored its clock on 97 snare matches. Tested on the kick lane, which neither clock was fitted on, that alignment recovers 226 of 504 kicks, or 44.8 percent, across 59.3 to 506.1 seconds. The metronome clock recovers 286 of 504, or 56.7 percent, across 41.3 to 528.5 seconds. It wins by 60 kicks while covering 41 more seconds of the song, so it is not buying accuracy with a shorter span. A cymbal lane is the densest and least periodic thing on the kit, which is why a clock error punished it hardest and left snare, kick and tom looking fine.
Two things this does NOT overturn. The tom result stands and in fact moved the other way, 70.2 down to 60.1 percent confirm while still separating at 4.51 times, and that is stated here rather than buried. And the 101-collision comparison below has now been rerun on the new clock and its conclusion survives: the 101 confirm at 36.6 percent against 40.4 for the other 1,472, a gap of -3.7 points at z = -0.74, so the collision criterion still selects nothing. Both rates roughly doubled with the better clock. Full rebuild: watermelon-metronome-clock.
The short version. The 101 cymbal strokes flagged for colliding with a tom confirm against the cymbal stem at 19.8 percent. Every other cymbal note in the tab confirms at 19.3 percent. The collision criterion selected nothing. The claim that followed, that the whole 1,573-note cymbal lane runs at only 1.46 times chance on its own stem, is withdrawn: on the metronome clock the same notes reach 3.17 times. The lane scores in that session were snare 6.89, kick 5.21 and toms 7.71 through identical machinery, and the cymbal gap against them was the clock, not the notation.
Corrected 2026-09-08 by the leakage gate. The sentence above about the lane running at 1.46 times chance is a statement about the DETECTOR, never about the notation. On the 865 cleanest cymbals in the song, those with no other drum within 50 ms, this detector recovers only 14.9 percent against a 13.3 percent floor, which is 1.12 times. It cannot see this instrument in this recording. Nothing about whether the cymbal notation is correct follows from the 1.46 figure, and no cymbal note can be removed on it.
A second result cuts against the repair list. All 49 sub-sixteenth tom positions sit a median of 15.4 ms from a tom-stem onset and 81.8 ms from the sixteenth grid, measured in bars where the clock is independently accurate to about 16 ms. Snapping them to sixteenths would move 49 notes away from the audio.
Stem event first, instrument identity second, score timing third, Guitar Pro lane last.
Every number below follows that order. No verdict on this page uses a hand-availability argument, and no verdict uses a chart lane to decide which drum was struck. The protocol was written to disk and hashed with none of the 101 yet scored, so the thresholds could not be tuned to the answer.
Pre-registration SHA-256 3c785296 7a669198 06a81dde 02212e50 76ad0737 98d74dbd 7ba43f12 51e63557
The alignment in hand was fitted on the original tab. This pass runs on the copy tab carrying your ear edits, so reusing that clock needed proof rather than an assumption. Both files hold 105 bars, 472 quarter notes, one tempo automation at 56 bpm, zero meter disagreements, and a maximum bar-start difference of 0.0000 ms. The clock is the same object in both files.
The map itself is anchored on 97 snare matches. Cymbals play no part in building it, which is what keeps this pass from grading the cymbal lane against a ruler the cymbal lane helped carve. Scoring is restricted to the anchor span, 59.33 s to 506.08 s, which is 82 percent of the song. Anything outside is marked not scored and never counted as absent.
Same detector, same clock, same tolerance, same one-to-one matching for every row. Chance is measured by drawing 20,000 random timestamps inside the span and running them through the identical procedure.
| Notated lane | Scored against | Notes | Stem onsets | Confirm | Chance | Ratio | Reading |
|---|---|---|---|---|---|---|---|
| Tom | toms stem | 168 | 867 | 0.702 | 0.091 | 7.71× | positive control, strong |
| Snare | snare stem | 342 | 659 | 0.477 | 0.069 | 6.89× | positive control, strong |
| Kick | kick stem | 504 | 860 | 0.448 | 0.086 | 5.21× | positive control, independent band |
| Snare | cymbals stem | 342 | 1286 | 0.272 | 0.133 | 2.05× | cross-lane leakage, expected |
| Cymbal | cymbals stem | 1573 | 1286 | 0.193 | 0.133 | 1.46× (superseded: 3.17×) | its own stem, barely above chance |
| Cymbal | snare stem | 1573 | 659 | 0.091 | 0.069 | 1.31× | near null |
| Cymbal | bass stem | 1573 | 2216 | 0.203 | 0.219 | 0.93× | negative control, correctly null |
Superseded: the cymbal figure in this paragraph is 3.17 on the metronome clock.
Three lanes clear five times chance on their own stems. The cymbal lane sits at 1.46 on its own stem and 0.93 on an unrelated one. This replicates the earlier two-way page through a stricter method, since one-to-one matching stops a single loud crash from confirming a run of written notes.
If the clock were smearing cymbal hits, widening the window would recover them. It does not. The ratio peaks at the pre-registered 25 ms and decays in both directions, which is the shape of a lane that has no signal to find rather than a lane that is merely mistimed.
| Tolerance | Confirm | Chance | Ratio | |
|---|---|---|---|---|
| 10 ms | 0.065 | 0.058 | 1.13× | |
| 15 ms | 0.121 | 0.086 | 1.40× | |
| 25 ms | 0.193 | 0.133 | 1.46× | |
| 40 ms | 0.260 | 0.183 | 1.42× | |
| 60 ms | 0.311 | 0.238 | 1.31× | |
| 80 ms | 0.353 | 0.298 | 1.18× | |
| 120 ms | 0.441 | 0.417 | 1.06× |
The clock was inherited and proved identical across both score files, and that says nothing about whether it maps onto this recording. The spec requires a deliberately incorrect recording that preserves nuisance structure and breaks the correspondence. Watermelon's notated events were scored, through the Watermelon clock, against Keep It Greasey, a different Zappa performance run through the same Moises separator, 501.84 s against 545.27 s.
| Notated lane | n | Right recording | Wrong recording | Separation | Verdict |
|---|---|---|---|---|---|
| Tom | 168 | 7.71× | 1.36× | 5.67× | separates |
| Snare | 342 | 6.89× | 0.78× | 8.86× | separates |
| Kick | 504 | 5.21× | 0.70× | 7.49× | separates |
| Cymbal | 1573 | 1.46× | 1.00× | 1.45× superseded: 3.00× | weak now SEPARATES |
The clock is doing real work. Three lanes collapse to 1.36, 0.78 and 0.70 against the wrong performance, from 7.71, 6.89 and 5.21 against the right one. An alignment that matched anything rhythmic would not behave that way, so the map is specific to this recording.
Superseded. The paragraph that follows reasons from the 1.45× separation, which the metronome clock replaces with 3.00×. Its conclusion that the lane cannot promote an individual edit no longer follows from the number it cites.
This sharpens the cymbal reading rather than killing it. The wrong recording returns the cymbal lane at exactly 1.00×, a clean null, and its own recording returns 1.46×. Some cymbal notation is genuinely supported by this stem. The separation of 1.45× sits under the 1.5× floor, so the lane characterises well and still cannot promote an individual edit. Written into data/wrong_recording.json.
The 101 are tom-coincident by construction. If a tom transient leaks into the cymbal stem, every one of them confirms for the wrong reason and the pass is worthless. So the control measures notated tom onsets that carry no notated cymbal within the tolerance, scored against the cymbal stem by the same procedure.
A lone tom hit produces a cymbal-stem confirmation less often than a random timestamp does. The separation is clean and the source separator earned its keep. The pass is measuring cymbals.
Reproduction first. Applying the audit's own rule to the baseline file returns 24 measures and 106 colliding slots. Dropping the three measures whose cymbal voice does not fill the bar removes exactly 5 slots and leaves 21 measures and 101 collisions, with the sub-sixteenth tom count landing on 41 independently. The set is the same set.
The comparison that decides it. The 101 confirm at 19.8 percent. The other 1,472 cymbal notes confirm at 19.3 percent. A difference of half a percentage point on 101 events is nothing. Colliding with a tom does not predict whether the cymbal stem carries the stroke, so the collision list was never a list of suspect notes. The list sampled an evenly weak lane.
Eleven toms with eleven coincident cymbal strokes, checked one event against one audio event.
| Slot | Cymbal written | Stem time | Verdict | Tom at that slot |
|---|---|---|---|---|
| 4 | Ride (middle) | 310.825 | ABSENT | High Tom (48) |
| 8 | Ride (middle) | 311.908 | ABSENT | High Tom (48) |
| 9 | Ride (middle) | 312.179 | CONFIRMED | High Tom (48) |
| 10 | Ride (middle) | 312.450 | ABSENT | High Tom (48) |
| 11 | Ride (middle) | 312.721 | CONFIRMED | High Tom (48) |
| 12 | Ride (middle) | 312.992 | ABSENT | High Tom (48) |
| 13 | Ride (middle) | 313.263 | ABSENT | High Tom (48) |
| 14 | Ride (middle) | 313.534 | ABSENT | High Tom (48) |
| 15 | Ride (middle) | 313.805 | ABSENT | High Tom (48) |
| 16 | Ride (middle) | 314.076 | CONFIRMED | High Tom (48) |
| 19 | Ride (middle) | 314.889 | ABSENT | Very Low Tom (43) |
Three of eleven confirm, which is 27 percent against a 13.3 percent floor on a sample of eleven. That is one or two events above chance and it settles nothing on its own. The bar is worth opening for the tom reason below, and the cymbal verdicts here stay in the queue.
With confirmation at 19.3 percent against a 13.3 percent floor, roughly a third of any single CONFIRMED is signal and the rest is coincidence. That is enough to characterise a lane of 1,573 notes and it is not enough to license deleting one note. Per-event cymbal verdicts on this song are investigation queue entries. The same machinery on toms, snare and kick clears five times chance, so per-event verdicts there would carry real weight.
What the page therefore refuses to do. No cymbal note is removed. No cymbal note is added. The 101 verdicts are published as a queue with their timestamps so any one of them can be opened against the audio. The earlier framing said a hand has to leave the cymbal, and that argument is retired here because the cymbal stem can answer the question directly and did.
The repair list proposed pulling 41 fractional tom positions onto the sixteenth grid. The tom stem scores 7.71 times chance, so it has the standing to arbitrate. All 49 fractional tom notes in the file were scored.
Those fractional positions were generated from raw stem onsets, so matching them back to stem onsets proves provenance and cannot by itself prove the drummer played off the grid. The control that breaks it is the tom notes written on the grid, which were never derived that way.
| Population | n | Median distance to nearest tom onset | Within 25 ms |
|---|---|---|---|
| Tom notes written ON the sixteenth grid | 119 | 16.4 ms | 65% |
| Tom notes written OFF the grid | 49 | 15.4 ms | 84% |
| Snare in the 17 bars holding off-grid toms | 61 | 16.9 ms | 64% |
| Snare in every other bar | 281 | 11.6 ms | 63% |
The clock lands within about 16 ms in exactly the bars under suspicion, measured two independent ways. The grid sits a median 68.1 ms from the fractional position, which is four times the clock's own error. Snapping those 49 notes to sixteenths would move each one roughly four clock-errors away from the onset the tom stem records. The regrid is contraindicated and the item belongs in the queue as a per-fill question about tuplets and odd meters.
Post-transient resonance peak between 60 and 400 Hz on the isolated tom stem, 20 ms after the attack over a 120 ms window. FFT bin resolution is 5.38 Hz, so differences under about two bins carry no weight. Ordering is the claim. Absolute drum identity is not.
| Slot | Written | MIDI | Resonance peak | Centroid | RMS |
|---|---|---|---|---|---|
| 12 | High Tom | 48 | 129.2 Hz | 153.1 | 0.0166 |
| 13 | High Tom | 48 | 129.2 Hz | 158.4 | 0.0143 |
| 14 | High Tom | 48 | 129.2 Hz | 161.0 | 0.0167 |
| 15 | Mid Tom | 47 | 96.9 Hz | 144.2 | 0.0127 |
| 16 | Mid Tom | 47 | 102.3 Hz | 150.3 | 0.0148 |
| 17 | Mid Tom | 47 | 86.1 Hz | 122.1 | 0.0146 |
| 18 | Mid Tom | 47 | 86.1 Hz | 132.7 | 0.0129 |
| 19 | Low Floor Tom | 41 | 80.7 Hz | 127.0 | 0.0116 |
19 of 19 pairwise orderings agree. The written descent matches the measured descent at every comparison, running 129.2 Hz down to 80.7 Hz. Your predicted figures were 129 to 133 Hz for the first event and 85 to 102 Hz for the second group, and the stem returns 129.2 and 86.1 to 102.3. The hand correction holds on audio evidence, with the direction certain and the exact drum names still open.
| Slot | Written | MIDI | Resonance peak | RMS | Reading |
|---|---|---|---|---|---|
| 4 | High Tom | 48 | 91.5 Hz | 0.0180 | one drum |
| 8 | High Tom | 48 | 86.1 Hz | 0.0166 | one drum |
| 9 | High Tom | 48 | 91.5 Hz | 0.0182 | one drum |
| 10 | High Tom | 48 | 91.5 Hz | 0.0163 | one drum |
| 11 | High Tom | 48 | 86.1 Hz | 0.0126 | decaying |
| 12 | High Tom | 48 | 86.1 Hz | 0.0089 | decaying |
| 13 | High Tom | 48 | 86.1 Hz | 0.0079 | decaying |
| 14 | High Tom | 48 | 86.1 Hz | 0.0070 | decaying |
| 15 | High Tom | 48 | 86.1 Hz | 0.0073 | decaying |
| 16 | High Tom | 48 | 86.1 Hz | 0.0067 | decaying |
| 19 | Very Low Tom | 43 | 113.0 Hz | 0.0174 | higher, and a fresh stroke |
WITHDRAWN 2026-09-08, same day. This bar is UNRESOLVED, and neither side of it is a finding.
The measurement is withdrawn on its own terms. I first read 0 of 10 order agreement and called the fill backwards, on a measured 113.0 Hz final stroke against 86 to 91 Hz for the ten before it. That claim does not survive its own instrument. The repeated 86.1 Hz across six consecutive strokes is a one-bin reading at 5.38 Hz resolution, which is the signature of a tracker locking onto a single resonance, and a separated tom stem makes that worse. A single-F0 read cannot support a direction claim under those conditions, and it certainly cannot resolve a stroke that may be more than one drum.
The chart cannot settle it either, and saying so is the point. I cropped chart bar 48 and read a shape there. That read is not reported as a finding. This scan measures 5.0 px per staff space against a 15 px floor for notehead work, already recorded as q-2026-09-06-c45c11, which marked Watermelon lane counts blocked on resolution and asked for a scan at roughly three times this one. A notehead count taken below that floor is exactly what the floor exists to reject, and my own crop does not get an exemption for being mine.
What actually survives on this bar. Ten strokes hold a flat resonance near 86 to 91 Hz with a clean rms decay from 0.0180 to 0.0067, which reads as one drum through a decrescendo. That is a loudness contour, and loudness does not depend on the F0 estimate that failed. The eleventh stroke is unresolved.
Standing state: bar 57 waits on a higher-resolution scan. No edit is proposed for it, no drum is named, and no notehead is counted from this page image.
Two runs reported 154 chart heads against 193 tab snares, and 124 chart heads against 340 tab events. Both are correct and they count different things on both sides.
| Side | Number | What it actually counts |
|---|---|---|
| Tab | 193 | Plain, non-ghost snare instances. Identical in both revisions. |
| Tab | 342 | All snare instances in r8968524, which is 193 plain plus 149 restored ghosts. The quoted 340 sits two off this. |
| Tab | 193 | Total in the pre-sweep original, which carries zero ghosts, so the two definitions collapse there. |
| Chart | 154 | Snare-lane heads from the erosion reader with the ink cut at 180 or above. |
| Chart | 83 / 136 / 137 | Three earlier reader passes held on this disk, same page images. |
| Chart | 124 | A parallel session's reader pass. Not present on this disk, so it cannot be re-derived here. |
Neither side holds the same quantity across the two comparisons. Until a comparison names its revision, its ghost rule, its bar range and its reader pass, the gap between chart and tab carries no information about what the drummer played. One chart image supports four different head counts on this machine alone, which is a reason to distrust the reader in dense passages and to re-read from the stem.
The governing spec arrived after this pass and is now installed as /perfect-stem-read. It requires a snare count to declare its revision, measure range, voices, ghost policy, grace policy, tie and duplicate handling, event definition and reader version before the number supports any claim. Declaring those fields surfaced something the earlier reconciliation missed.
| Field | PRESWEEP r7715683 | BASELINE r8968524 |
|---|---|---|
| Artifact SHA-256, first 16 | 687445e3371131fd | 4376b4a39dd1a466 |
| Track | 8, Vinnie Colaiuta | 8, Vinnie Colaiuta |
| Measure range | all 105 bars | all 105 bars |
| Voices | [0] | [0, 1] |
| Event definition | note instance, never a Note definition id | |
| MIDI counted | 40 | 40 |
| Total instances | 193 | 342 |
| Plain / ghost | 193 / 0 | 193 / 149 |
| Grace notes | 0 | 0 |
| Distinct sixteenth positions | 193 | 293 |
| Positions carrying two notes | 0 | 49 |
The 342 occupy 293 positions, and all 49 doubled positions share one pattern. Every one is a plain snare in voice 0 stacked with a ghost snare in voice 1 at the identical bar and sixteenth slot. The pattern is unanimous across all 49, with no other combination appearing.
A drummer cannot strike one snare loudly and quietly at the same instant. These are 49 notation collisions the ghost restore introduced, and they are a separate defect from anything the cymbal pass measured. Written into data/snare_duplicates.json and data/snare_scope_full.json. No edit is proposed here, since the restore's provenance decides this rather than audio.
The spec requires more than one bleed control: target and competing transient energy, target and competing template similarity, and the detector's firing rate through known target-silent passages, all calibrated from confirmed events inside this recording. Running it changed what the cymbal figures mean.
An isolated event is one with no other drum notated within 50 ms. That gives 865 isolated cymbals and 59 isolated toms. The cymbal template is what the cymbal stem looks like on an isolated cymbal. The bleed template is what the same stem looks like when a tom sounds and no cymbal is written. The margin below is similarity to the cymbal template minus similarity to the bleed template.
| Population | n | Margin median | Cymbal-stem energy | Tom-stem energy |
|---|---|---|---|---|
| Isolated cymbal (target) | 400 | +0.0078 | 168.1 | 0.2 |
| Isolated tom (bleed signature) | 59 | −0.0071 | 87.6 | 182.7 |
| The 101, all together | 101 | −0.0038 | 148.2 | 154.4 |
A tom alone puts 87.6 of energy into the cymbal stem against 168.1 for a real isolated cymbal, so this separation leaks roughly half a cymbal's worth of energy on every tom hit. That is why the gate was necessary.
| Subset of the 101 | n | Margin p25 | Median | p75 | Reads as |
|---|---|---|---|---|---|
| Onset CONFIRMED | 20 | −0.0051 | +0.0131 | +0.0148 | cymbal, above the isolated-cymbal median |
| Onset ABSENT | 80 | −0.0125 | −0.0058 | +0.0008 | near the tom-bleed median |
Mann-Whitney, CONFIRMED above ABSENT, p = 0.0001. Onset detection and spectral shape are independent measurements and they agree. The 20 confirmed strokes are real cymbal, supported twice.
The cleanest cymbals score WORSE than the tom-coincident ones. Isolated cymbals reach 1.12×, the whole lane 1.46×, the 101 1.49×. A detector that finds genuine cymbals barely above chance in the easiest possible conditions is not measuring the notation. It is failing to see a ride pattern inside its own resonant bed, which is the exact difficulty the spec names for broadband cymbal textures.
So ABSENT carries no information and never licenses a removal. Eighty-five percent of cymbals that are known good by isolation come back ABSENT. Applying that verdict as a repair instruction would delete most of a correctly notated lane. Every one of the 101 stays.
Only four stretches run 1.5 s or longer with no notated cymbal, totalling 19.4 s, because the notation is nearly continuous. The detector fires at 0.98/s inside them against 2.96/s elsewhere, a ratio of 0.33. Notation and stem agree at that coarse level, and the sample is four passages, so it carries little weight on its own.
The recall failure above could be a tuning failure, so it was worth attacking directly. Five onset functions were tried on the cymbal stem: plain spectral flux, SuperFlux (flux against a frequency-max-filtered lagged frame, which suppresses a ringing decay), adaptive bed subtraction, complex-domain deviation and its rectified form. Phase-aware methods were included because a fresh stick attack breaks phase continuity even while the cymbal keeps ringing. Parameters were tuned on 432 odd-indexed exemplars and reported on the 433 even-indexed holdout.
| Detector | Holdout recall | Chance | Ratio |
|---|---|---|---|
| Baseline, plain flux 4-16 kHz | 0.263 | 0.220 | 1.20× |
| SuperFlux on a bed-subtracted spectrum | 0.199 | 0.124 | 1.60× |
| Complex-domain deviation topped out at 1.72× on tune, SuperFlux alone at 1.65×, bed-residual energy at 0.80×. | |||
A real 1.34× improvement on held-out data, and the 3× target is not met. Five independent onset functions across five frequency bands and every threshold all land between 1.2× and 1.7×. That flatness is the result, since a tuning problem does not resist five different methods equally.
| Global shift | Cymbal ratio | Snare ratio (control) |
|---|---|---|
| −100 ms | 0.73× | 3.07× |
| −50 ms | 0.99× | 4.85× |
| 0 ms | 1.64× | 6.89× |
| +50 ms | 0.73× | 4.59× |
| +100 ms | 0.81× | 3.26× |
| +275 ms | 2.20× (its best) | — |
The snare peaks cleanly at zero and falls off symmetrically. That is what a correct alignment looks like, and it confirms the clock once more.
The cymbal lane has no peak at zero. Its best bump sits at +275 ms, which is 1.03 sixteenth notes at this tempo, with bumps of similar height at −250 and +300. 93.0 percent of written cymbal intervals fall within 10 percent of one sixteenth, median 271.2 ms against a 267.9 ms sixteenth.
A lane written on nearly every sixteenth matches itself one sixteenth over, so signal and null have the same shape and no offset is identifiable. Onset matching cannot validate a continuous sixteenth-note ride. That is a property of the method against this material, never a verdict on the notation, and it is why the snare works at 6.89× while the cymbal cannot be scored at all. Written into data/ridebed_final.json and data/ridebed3.json.
Per-event matching is unidentifiable on this lane, so the question changed. Ask whether the ride's periodicity matches the stem, over 6-second windows stepped by 2 seconds, with the same Keep It Greasey wrong recording as control.
| Recording | Median lag / notated spacing | Windows within 15% of 1.0 | Median autocorrelation |
|---|---|---|---|
| Watermelon (right) | 1.971 | 36 of 216 | 0.217 |
| Keep It Greasey (control) | 0.814 | 15 of 216 | 0.252 |
The stem's dominant pulse sits at almost exactly twice the notated ride spacing, which is 271.0 ms, so the envelope pulses at the eighth while the tab writes sixteenths. Separation against the wrong recording is 2.40×.
If the drummer played eighths and the tab wrote sixteenths, the off-eighth notes would be inventions. Splitting purely on score position, with no reference to audio, refutes that.
| Position | n | Confirmed | Recall | Ratio |
|---|---|---|---|---|
| On the eighth | 785 | 211 | 211/785 = 0.269 | 2.17× |
| Off-eighth sixteenth | 736 | 151 | 151/736 = 0.205 | 1.66× |
Only 1.31× apart, Fisher p = 0.0021. Real and far too small for fabrication. The off-eighth sixteenths carry genuine support at 1.66×, so the drummer played the sixteenths.
A sixteenth ride with every other stroke accented pulses at the eighth in an envelope and still puts a real attack on every sixteenth. Both results then agree. The pre-sweep tab r7715683 encodes exactly that, marking 1,317 of its 1,778 ride notes with AntiAccent. Splitting on that flag alone, with no reference to audio:
| Author's marking | n | Confirmed | Recall | Ratio |
|---|---|---|---|---|
| PLAIN, the author's accents | 416 | 122 | 122/416 = 0.293 | 2.37× |
| GHOSTED, the author's weak strokes | 1185 | 240 | 240/1185 = 0.203 | 1.64× |
1.45× apart, Fisher p = 0.00012. The notes the transcriber marked as accented are measurably louder in the isolated cymbal stem than the ones marked weak, on a split made from a notation flag with the audio never consulted.
Three independent splits, all significant, all in the same direction: envelope periodicity at 1.971, score position at 1.31× (p = 0.0021), author's flag at 1.45× (p = 0.00012). The modest effect sizes are what a detector with 14.9 percent recall can deliver, since poor recall caps how far any split can separate.
The consequence for the tab. Those 1,317 ride ghost flags encode an accent structure the audio independently supports, and the current copy tab r8968524 carries zero cymbal ghosts against the pre-sweep tab's 1,317. The accent grid has been lost from the working file. This is a provenance repair with audio backing, and it is the strongest actionable result on this page. Data: beat_periodicity.json, eighths_test.json, accent_grid.json, g44_flag_parity.json.