Müller and the Watermelon Repairs

Three Meinard Müller texts, the drum-tab forensics on Watermelon In Easter Hay (s35881), and what happened when the replacement methods were actually run against the recording. Measured 2026-09-06.

1. Two kinds of claim, kept apart

An earlier version of this page said every hand-rolled step "failed for a reason the texts state." That sentence merged two different facts. A textbook explaining why a procedure is vulnerable is one thing. A rerun showing the replacement changes the Watermelon verdict is another. Every row below is now labelled with which one it is.

2. What was measured

2.1 Independent census of the tab

MEASURED The .gp is a Guitar Pro 7 zip, so PyGuitarPro cannot read it. A GPIF XML parser written this session reads track 8, Vinnie Colaiuta, and reproduces the handoff's census without consulting it.

2.2 The formal landmark test, Gate N9

MEASURED Brandon's four chicks are audio timestamps, so they need no alignment. Detector pre-registered before scoring: HPSS-percussive, γ=100, M=10, 0.04 s refractory, density target 2.0/s, tolerance ±50 ms.

hits               2 of 4  (50%)
residuals (ms)     [-35.4, -34.1]
FAILED landmarks   [181.93, 183.03]
nearest onset (ms) [71.3, 80.0]        a miss is a distance, not a silence
pattern-rotation   p_perm = 0.2020 over 2000 rotations (null mean 0.78 hits)
coverage           p_cover = 0.1988, expected false 0.80 of 4
binomial           p = 0.179   ADVISORY ONLY, assumes independent draws
VERDICT            UNDETERMINED
  - only 4 landmarks, under the minimum 8
  - hit coverage 50%, under the required 75%
  - pattern-rotation p_perm = 0.2020, over 0.01
  - no selection rule stated, so the set is not reproducible

The binomial is now reported as advisory only. Consecutive pulses are correlated through meter, so independent Bernoulli trials are not licensed. The governing figure is p_perm, obtained by rotating the whole four-landmark pattern with its inter-landmark intervals intact.

2.3 Detector density sweep, pre-registered

MEASURED The full sweep is reported so that no density can be selected after seeing which one flatters the result. Adding onsets buys a hit and costs more coverage, so the verdict never moves.

The same two landmarks persist across the whole range with residuals stable at −35.4 and −34.1 ms. That persistence is evidence about those two instants. It says nothing about the other roughly 170.

2.4 The single affine map, measured rather than assumed

MEASURED An earlier draft asserted that a rubato ballad cannot admit one offset-and-scale map. That was a claim, so it was tested. The old two-parameter model was reimplemented, fitted by grid search on the sparse 193-event snare lane, and then scored on its residual structure rather than on its headline percentage.

transform          offset -1.750 s, scale 1.01400     (handoff reported scale 1.012)
one-to-one matched 79 of 193 = 40.9% of notated, 7.2% of detected
unmatched notated  114 (59.1%)      unmatched detected 1011 (92.8%)
COLLAPSE DIAG      busiest onset 1, attacks/used onset 1.00     <- clean, no pile-up
residual median    -2.9 ms,  |median| 24.7 ms, IQR 12.9-36.6 ms
DRIFT slope        -52.4 ms per 1000 s  =>  -23.2 ms across the matched span
DRIFT Spearman     rho = -0.310
-> AFFINE ASSUMPTION SUSPECT: residuals walk with score position

The section table is where the model visibly comes apart. A global 40.9% conceals a match rate that swings by a factor of twenty across the piece.

The collapse diagnostic is clean at 1.00 attacks per used onset, so this is not a pile-up artifact. The model fits two adjacent sections at 91.7% and 83.3% and then falls to 8.7%. That is the signature of a performance whose tempo the map cannot follow, and it is now a measurement on Watermelon rather than a claim about ballads in general. This is the strongest empirical result on the page.

2.5 The printed chart, and what it did to my own experiment

MEASURED The five Drumnet chart pages are on this Mac and had not been read. Page 1 carries three facts that bear directly on everything above.

The search was rerun with the range the chart implies, offset 0 to 130 s. Result: offset 66.700 s, scale 1.0140, 81 of 193 = 42.0%, against 40.9% at offset −1.750 s. Two alignments sixty-eight seconds apart score within 1.1 points of each other.

2.6 Eight same-album controls, identical pipeline

MEASURED Same detector, same tolerance, same grid, same landmark set, no per-recording retuning. A control is a recording the Watermelon tab does not transcribe, so a high score there is a false positive.

2.7 Does the chart notate a foot hi-hat?

INCONCLUSIVE AT THIS SCAN QUALITY This is step 2 of the evidentiary order and the question that would actually move Repair 1. At 7x magnification on page 1 the chart shows x-noteheads above the staff, kick noteheads in the bottom space, accent marks below the staff, and at least one x-notehead inside the staff. I could not reliably resolve whether any notehead sits in the first space below the staff, which is where Weinberg places the foot hi-hat.

Source for that reading: ~/Projects/_outputs/zappa-drum-sources/06-drumnet/Watermelon-In-Easter-Hay_p1.jpg, 595 × 842 px, 70,150 bytes, cropped to the first two systems and resampled 7x with Lanczos. The scan resolution is the limiting factor, so a higher-resolution source is what this needs.

All five pages need a careful reading at higher resolution before this question is answered. A chart that omits the texture did not transcribe it, and that is never evidence the record lacks it.

2.8 The result that reframes the lane

2.9 The DP replacement was run, and it did not fix the problem

MEASURED align_dp.py, the IRMM §5.3 anchor-partitioned DP over onset sequences with a tempo-invariant log-IOI local cost, arrived from a parallel session and was run on the same data. Unanchored:

The wrong recording wins again, and residuals run to hundreds of seconds. A log-IOI cost is tempo-invariant, which also makes it position-invariant, so without anchors it matches rhythmic shape anywhere in the timeline. Replacing the alignment method did not restore discriminating power. Anchors are the missing ingredient, and §2.8 is why supplying them is not yet possible.

2.95 RESOLVED, and it corrects a conclusion above

Run through the gate against a block null, the handoff transform still returns UNDETERMINED: recall 31.6% against a 21.2% null, z = 2.85, gain +10.5 points, p_rank = 0.013, under the z ≥ 5.0 bar. Consistent with the full mix being the wrong instrument for a snare lane. The alignment is now trustworthy; the detector is the limit.

2.10 A safety finding that halts any .gp write

2.99 Cross-Mac results, 2026-09-06 evening

The affine measure fails on an isolated stem too

Isolating the instrument raised the rate 3.7 points and moved the winning offset by 40 seconds. A wrong recording still lands within 2.1 points of the right one. The non-discriminating result holds on a stem exactly as on a mix, which makes q-2026-09-06-57b3e8, the onset-based DP alignment, the binding item, since every downstream test runs through that map.

2.995 The anchored DP alignment works

2.999 The lane-label trap, and the strongest audio evidence yet

2.9999 DECISIVE: the 4-of-4 does not survive a stated selection rule

3. The method table, honestly labelled

we prefer to have missing matches over having bad or wrong matches. Hence, we do not force the alignment of all score notes but rather allow note objects to remain unmatched. Müller, Information Retrieval for Music and Motion, §5.3 introduction, printed p97. The DP that implements it is §5.3.3, printed p99-102.

4. Notation authority is not performance evidence

The evidentiary order for any repair

1. Does the audio carry the event at that instant? Gate N9, with a stated selection rule and a correlation-respecting null.
2. Does an independent printed witness notate it? The five Drumnet chart pages are the untried source here.
3. Does the tab omit an event that 1 or 2 support? Only survivors are candidates.
4. How should a surviving event be written? Weinberg applies at this step and no earlier.

A repair is built per event, from events that passed. "Roughly 172 events" is a hypothesis and never a work order.

5. The gate, and what changed in it

~/.claude/skills/notation-evidence-gate/verdict_gate.py. Folded into the existing skill rather than shipped as a competitor.

• Seven quantities reported separately: precision, recall, F, signed residual bias, residual spread, unmatched-notated rate, unmatched-detector rate. Maximum pre-matching occupancy is printed as a collapse diagnostic only and is no longer part of any score, since one-to-one matching already removed many-to-one assignment.
• Null admissibility is checked first. Onset-density CV across blocks is measured, and a global rotation is refused above 0.60. Block rotation, which moves each event only inside its own window, is the default.
• Counts mode can no longer confirm. Without event times it cannot build a correlation-respecting null, and a binomial over correlated musical events is not licensed. It can return UNUSABLE or UNDETERMINED only.
• Gate N9 requires a stated selection rule, at least 8 landmarks, 75% hit coverage, a pattern-rotation p, and an explicit list of failed landmarks with their nearest-onset distance.

6. Not done, queued rather than dropped

Also open: the −35 ms residual, the 39.56 s score-to-audio duration difference, and the five Drumnet chart pages, which are the printed witness step 2 needs and have not been read.