Watermelon: the isolated-stem rerun

2 / 4hat-stem landmark hits

0.1995p_perm, hat stem

0.202p_perm, full mix

44.6%snare-stem affine rate

42.5%rate of a WRONG recording

−49.4 msmean hat-stem lead

4 / 4leave-one-out on the bias

545.27 sstem and source agree

The short version

REPAIR 1 STAYS UNAUTHORISED The isolated stem changes nothing about the landmark verdict. Full mix returned 2 of 4 at p_perm 0.202. The hat stem returns 2 of 4 at p_perm 0.1995.

AFFINE MEASURE FAILS ON A STEM TOO The snare stem scores 86 of 193 at offset 38.400 s, against 79 of 193 at offset −1.750 s on the mix. Two fits of one recording landing 40 seconds apart, and 44.6% barely clears the 42.5% a recording the tab never transcribes already scored.

THE RESIDUAL IS DIAGNOSED All four chick landmarks sit on hat-stem onsets that arrive early: mean −49.4 ms, median −52.7 ms, spread 25.1 ms. Leave-one-out holds 4 of 4. This is the open −35 ms item, and it is a constant lead rather than scatter.

On this page

Preconditions
Detector settings
Gate N9, hat stem
Snare control
Affine alignment
The residual
What this does not license
Correction to the skill
Reproduce
Queue

1. Preconditions, verified before any measurement

Check	Result
Stems present on this Mac	15 files, 96 MB each
Stem duration	545.266667 s on every stem
Source FLAC duration	545.27 s
Time transfer	durations agree, so audio timestamps carry across unchanged
Tab census reproduced here	193 snare, 1,724 ride, 107 kick, 6 pedal hi-hat, 2,091 total
Weinberg reference authenticated	SHA256 7e9288210467177f565b367a3c6a3bd36a3e7bcbb75c6eab28325a82bf30ee12, 256,197 bytes, identical to the public article

The 112bpm-442hz in the folder name is detected metadata. Nothing was time-stretched, so the iMac's landmark times apply without correction.

2. Detector, pre-registered parameters, no retuning

HPSS-percussive, gamma 100, M 10, refractory 0.04 s, density target 2.0 per second, tolerance 0.05 s. Identical to run_controls.sh.

Stem	Onsets	Threshold	Density	Duration
snare	1,090	0.00370	1.999/s	545.27 s
hat	1,091	0.00085	2.001/s	545.27 s

3. Gate N9 landmark test, hat stem

Landmarks 181.93, 183.03, 184.10, 185.19 s.

hits               2 of 4  (50%)
residuals (ms)     [-48.1, -33.5]
median |residual|  40.8 ms
FAILED landmarks   [184.1, 185.19]
nearest onset (ms) [58.6, 57.3]
pattern-rotation   p_perm = 0.1995 over 2000 rotations (mean 0.76 hits)
coverage           p_cover = 0.1951, expected false 0.78 of 4
binomial           p = 0.173  (advisory only)
VERDICT            UNDETERMINED

Four reasons the gate refuses it: only 4 landmarks against a minimum of 8, hit coverage 50% against a required 75%, p_perm 0.1995 over the 0.01 bar, and no stated selection rule.

4. Negative control, same landmarks against the snare stem

hits               1 of 4  (25%)
FAILED landmarks   [181.93, 183.03, 185.19]
nearest onset (ms) [71.3, 683.7, 150.2]
pattern-rotation   p_perm = 0.5105
VERDICT            UNDETERMINED

The hat stem carries more of this pattern than the snare, and it still falls short.

5. Affine alignment, snare stem against the 193-event notated snare lane

Source	Offset	Scale	Hits	Rate
Snare stem, this Mac	38.400 s	1.0100	86 / 193	44.6%
Full mix, Jacks-iMac	−1.750 s	1.0140	79 / 193	40.9%
Full mix, wide window	66.700 s	1.0140	81 / 193	42.0%
102 Joe's Garage, a recording the tab does not transcribe	−1.350 s	1.2680	82 / 193	42.5%
Eight-control mean	—	—	—	26.1%

Isolating the instrument raised the rate by 3.7 points and moved the winning offset 40 seconds. A wrong recording still scores within 2.1 points of the right one. The measure has no discriminating power on this material, on a stem exactly as on a mix.

6. The residual, diagnosed

Signed nearest-onset distance, landmark to stem. Negative means the onset arrives early.

Landmark	Hat stem	Snare stem
181.93	−48.1 ms	−71.3 ms
183.03	−33.5 ms	−683.7 ms
184.10	−58.6 ms	−35.4 ms
185.19	−57.3 ms	−150.2 ms

Hat stem: mean −49.4 ms, median −52.7 ms, sd 11.6 ms, spread 25.1 ms, all four negative. Removing that constant bias, a fixed shift rather than a widened tolerance, puts 4 of 4 inside 50 ms.

Held out	Shift from the other three	Residual	Result
181.93	−57.3 ms	+9.2 ms	hit
183.03	−57.3 ms	+23.8 ms	hit
184.10	−48.1 ms	−10.5 ms	hit
185.19	−48.1 ms	−9.2 ms	hit

Landmark spacing averages 1.0867 s against the GP quarter of 1.0714 s at 56 BPM, a ratio of 1.0142 that matches both alignment scales found independently.

7. What this does not license

Gate N9 still fails. Four landmarks, no selection rule. A repair needs eight and a reproducible rule for choosing them.
The shift is fitted on the same four instants it then improves. Leave-one-out is reassuring across those four and it is not an out-of-window test.
The bias cannot be checked against the six pedal hi-hat events already in the tab. Carrying score time to audio time needs the affine map, and that map is discredited. Mapped through either fit, those six give residuals scattered from −242 ms to +980 ms, which measures the map rather than the texture.
A detector never outranks a human reading. The five Drumnet chart pages are the printed witness and they are still unread at high resolution.

8. Correction to the shipped skill

SKILL.md ruling 3 states that the four chicks appear in the hat stem at 181.92, 183.00, 184.09 and 185.17 s, which reads as 4 of 4. Run through verdict_gate.py with the pre-registered parameters, the hat stem gives 2 of 4, with the other two missing by 58.6 ms and 57.3 ms. The 4 of 4 reading is recoverable only after removing the constant −52.7 ms lead, which the ruling never mentions.

9. Reproduce

D=~/.claude/skills/notation-evidence-gate
W=~/Projects/_outputs/notation-evidence-gate/watermelon-stem-rerun-2026-09-06
PY=~/.venvs/notation-gate/bin/python
S="/Users/Shared/206 Watermelon in Easter Hay-E major-112bpm-442hz"

$PY $D/watermelon-evidence/novelty_onsets.py \
    "$S/206 Watermelon in Easter Hay-hat-E major-112bpm-442hz.wav" \
    --density 2.0 --out $W/ons_hat.txt

$PY $D/verdict_gate.py --landmarks 181.93,183.03,184.10,185.19 \
    --detected $W/ons_hat.txt --duration 545.27 --tolerance 0.05 \
    --label "4 chicks vs HAT STEM"

10. Queue

q-2026-09-06-855491 closed by this page, with the answer UNDETERMINED on better data.
q-2026-09-06-57b3e8 onset-based DP alignment, IRMM 5.3, open and now the priority, since the affine map is the binding constraint on every downstream test.
q-2026-09-06-083302 log-compressed novelty against the snare stem, partly served here.
q-2026-09-06-bcf816 score-informed NMF for per-note ride dynamics, open.

Run on MacBookPro, 2026-09-06. Skill notation-evidence-gate, bundle from Jacks-iMac via iCloud sfg/, installed as the union of skill-sync-2026-09-06/ and skill-sync-2026-09-06-notation-evidence-gate/. Working files at ~/Projects/_outputs/notation-evidence-gate/watermelon-stem-rerun-2026-09-06/. Notation authority: Norman Weinberg, "Guidelines for Drumset Notation", Percussive Notes, June 1994, drumnotes.com/docs/notation.pdf.