Watermelon In Easter Hay: counting the ride

A relayed detector design said to abort and switch separators if held-out error stayed near four notes a bar. Error fell to 0.84, positions landed at a median 24.6 ms from the written grid, and one criterion refused to move: the stem played backwards still fires almost as often as the stem played forwards. No ride note is written to the tab.

Songsterr s6857183, Vinnie Colaiuta drum staff, revision r8973454. Drum work only, no guitar transcription. Measured 2026-09-09 on MacBookPro against the Moises 15-stem set.

1The abort condition was wrong, and testing it was the whole job

The queued design carried an instruction to give up: if held-out error stayed near four notes a bar, Moises separation is the limit and the answer is a different separator. That instruction needed its own measurement, because a counting failure and an evidence failure look identical from outside.

The test asks one question no detector can beat. Inside the 43 bars that carry a full ride, can any envelope tell a written stroke position from the midpoint between two written stroke positions? Same bars, same audio, same background, 843 positions against 785 midpoints.

2Stage one of the design is inert on this stem

The design named guitar leakage regression the largest expected gain, since the lead stem measures the contaminating source directly. Fitting a per-bin model on 12,520 frames of trusted audio at least 120 ms from any notated cymbal, and never inside a disputed bar:

The premise was reasonable, because the bleed is real. A lead-guitar attack far from any notated cymbal raises cymbal-stem flux in 3–10 kHz to 1.75x the random-frame median at z 5.4. A notated ride stroke reaches 2.48x. Guitar bleed therefore arrives at 70.7% of the amplitude of a real ride stroke.

The design's own fallback was taken instead. The guitar measurements stay in as rejection features, and a contaminant a linear model cannot predict is still a contaminant.

3Eight stages, one dropped, one rebuilt

4The defect was my proposal geometry, not the stem

The first rebuild returned 3.35 notes a bar and a candidate-level held-out AUC of 0.594, near chance, which read as confirmation of the abort condition. Separability had already ruled that out, so the fault was mine.

Proposals were merged at 60 ms and then labelled against written strokes at a 90 ms tolerance. A sixteenth at 53–59 bpm is about 270 ms. Proposals landed every 82 ms, so neighbouring candidates were interchangeable and the labels were close to arbitrary. Sweeping both, with the tolerance constrained under half the merge:

0.16 s is taken over 0.19 s at equal error, because 0.19 puts the threshold back on a boundary. Boundary pinning is the exact defect this pass exists to remove.

5The acceptance gate, all five criteria

43 bars, 843 written strokes, both re-derived from the published Guitar Pro file rather than quoted. Group cross-validation splits whole bars. Refractory 90 ms.

6Two controls the gate never asked for

Five passes out of five is where a done-claim usually goes. Both extra controls said the gate measures less than it appears to, and chasing them is where the rest of this page comes from.

Only two thirds of detections sat inside 40 ms

The stem played backwards fires just as often

Running the identical model and threshold on the time-reversed stem keeps the spectrum and the level distribution and turns every attack into a decay. It returns 889 detections against 846 forward, a ratio of 0.95x.

7The 40 ms window was measuring the clock

The 40 ms target was written before the offset distribution was known. Measuring where detections actually land, against a circularly shifted control with the same count and spacing:

Median offset −11.8 ms, median absolute offset 24.6 ms, interquartile range −33 to +12 ms. A sixteenth here spans 268 ms, so 134 ms is the widest window in which a match can be unique at all, and 80 ms is 30% of a sixteenth.

8Two attempts at direction sensitivity, both measured, neither enough

Energy asymmetry alone did nothing at all. Reversed hard negatives moved it, and it stays far under the 3x target. A second criterion broke while fixing the first: at merge 0.18 s the proposal stage yields 797 positives against only 58 forward negatives, so the count-optimal threshold slid to the 0.0% boundary, which is the defect this whole line of work exists to remove.

9The disputed bars, and where the eye overrules the detector

No ride note is written to any of them. Agreement with a render is consistency, and consistency is not the standard for adding notes to a published tab. The matched-share half of that gate is met at a defensible 80 ms. The direction half was recorded as an adjudicated blocker rather than waved through, and section 12 settles it: the 3x target is retired and the paired counterfactual passes at 81.2%. The refusal to add notes stands on section 16 instead.

10Two provenance defects, both now repaired

11What is on disk

12The 3x gate is retired, and the detector passes what replaces it

A detection count on reversed audio measures how many energy events reversed audio contains, and reversed cymbal audio carries about twenty per bar. So 1.30x established nothing about attack morphology, and 3x would have established nothing either. The replacement asks the narrow question instead.

797 verified strokes, group cross-validation splitting whole bars, sign-flip permutation over 20,000 draws so the p-value floor is 5.0e-05. A0's columns are identical in both arms by construction, so its 49.3% is a wiring check on the harness and says nothing about the audio.

The detector is direction sensitive and always was. A1 is the representation that produced 1.30x, and on a fair paired test it prefers the forward ordering 81.2% of the time. The order-preserving columns multiply the margin by 13 and leave the proportion alone. The shallow nonlinear model scores lower on direction than the linear model on identical columns, which places the gain in the representation rather than in interactions among features.

13Two confounds caught before any of that was reported

Peak position was worth ten points

Proposals are peak-picked on the flux function, and the envelope peak lands a median +1.0 frames away from it. Time-reversing then moves the peak by 11.6 ms, so a template can score the reversed arm lower for a position reason. Recentring every patch on its own envelope peak, which puts the peak dead centre in 82.2% of patches, drops A1 from 91.1% to 81.2% and the median margin from +0.0301 to +0.0063. The 81.2% is the figure that survives.

My own bug: reshape returns a view

Two of my scripts reported 81.2% and 67.1% for one quantity. P.reshape(len(P), -1) returns a view, so V /= norm wrote unit-normalised values back into the patch array itself, and every later read of its energy was reading unit-norm patches. The script that used V = V / norm allocates and was correct throughout. After the fix both return 81.2%. Two of my own scripts disagreeing on one number is a bug report about me.

14The count gate was measuring the proposal grid

At merge 0.18 s the proposal stage returns 855 candidates in 43 bars whose written total is 843. Section 8 recorded the symptom, 797 positives against 58 forward negatives with the threshold on a boundary. The cause is worse than a threshold problem.

15One-to-one matching, and the gap that was computed rather than measured

The 80 ms window rested on two unmeasured assumptions: that matching was one-to-one, and that 268 ms was the minimum written ride spacing. Both are now measured. Across 800 within-bar gaps the minimum is 254.2 ms, with zero gaps under 200 ms, so the widest window in which a match can be unique is 127.1 ms rather than the 134 ms carried before. 80 ms sits below both figures.

The two matchings are identical through 100 ms and separate only at 127 ms, so the position result survives its caveat. The whole curve is published here and 80 ms is one named operating point on it.

16Bars 17 and 51, decided

Every bar number in this section indexes TEMPOFIX-INT-on-r8973389-Brandon-edit.gp track 8, the drum staff of Songsterr s6857183 revision r8973454, read through data/labels_gp.json.

Bars 17 and 51 write zero ride while the stem shows a ride at neighbour amplitude. Bars 99 to 105 were read as silent by eye, one bar per panel on a shared dB scale, and that made the coda the control this decision used. Bar numbers here are 1-based.

Question bars average 19.3 detections and coda bars average 19.9, a separation of 0.97x. That whole-block comparison is the weaker half now that bars 99 to 104 are known to carry some cymbal activity. The strong case is bar 105, which keeps 30 of 30 proposals, the most the detector keeps anywhere in the song, on 0.0070x of a stroke's energy, the least anywhere in the song. On that bar the score carries no information about whether the bar contains playing.

Amplitude separates them, and ring-out is refuted

Question bars sit at 0.609 to 0.803 of a verified stroke and every coda bar sits at 0.084 to 0.218, a 2.8x gap with no overlap. Both question bars follow a full-ride bar and a ride cymbal rings, so sustained energy needed its own test. Energy by quarter of the bar, normalised to each bar's own first quarter, gives Q4/Q1 of 1.07 for bar 17 and 0.82 for bar 51 against a full-ride median of 0.75. A decaying bar falls well below 1.0. Bar 87 gives 0.04, so it collapses by its last quarter and reads as the half-bar pause it was called.

17The three follow-ups, and the identity correction

The count claim is not earned at any density. Lowering the peak-pick separation admits more acoustic peaks with no lattice imposed. At 1.01 candidates per written note the classifier gives MAE 0.79 against accept-all's 0.84. At 3.26x it gives 4.28, at 4.03x 5.07, at 6.01x 5.21, against a target under 1.00, with thresholds mid-range at 60 to 67% of their searched range. The classifier only ties accept-all where it has nothing to reject.

The 149 direction failures are an alignment population. Peak offset from the flux pick separates 9.3% in its bottom quartile against 43.6% in its top, a 34.3 point spread, medians 5.805 ms against 46.440 ms. Waveform symmetry separates zero points, 19.5 against 19.5, so the obvious acoustic hypothesis is rejected by its own data.

Bars 17 and 51 were inspected directly, one bar per panel at native scale. Each carries about twenty sharp attacks, peaking at 0.748x and 0.628x of written full-ride bar 18, with HF share 58.6% and 56.9% against that control's 45.0%. Positions stay undetermined because the on-grid test gives coda control bar 104 87.5% against the written control's 68.8%, so it has no discriminative power.