Camarillo Brillo, read against the stems

1. What the five instants are

Parsed straight from s187082 r5563501 part index 3, the drum staff. Every one sits on beat four and carries the same three hand surfaces.

The count matches the queue item exactly, and the surfaces match. That part of the earlier sweep was accurate.

2. The assets, hashed before anything was measured

Fourteen stems, one folder, every file a distinct SHA256 and a decodable duration. No cloud placeholder, no truncated proxy, no repeated file standing in for a missing lane.

3. The recording identity gate

Before any timing-dependent claim, the score and the audio have to be the same performance. Three numbers disagreed straight away.

At the measured tempo this recording holds about 119 bars of 4/4. The tab claims 150. A bar count cannot be stretched by a quarter and stay the same performance.

Alignment, tested four ways with a positive control

A method that fails on everything proves nothing, so the same scorer was handed a train of onsets detected from the stems themselves. It had to pass before any failure verdict counted.

The pipeline detects a true alignment at z above twenty-three. It cannot find one for s187082 at any tempo, under Songsterr's own sync, or under a free affine search.

The second method agreed, and it was blunt about it

Dynamic time warping between the tab's rendered lanes and the stems' lanes, cost per step, lower is better.

Warp slope deviation ran to 2.5 where a real map sits near zero. One precision note on the reversal control, raised by an external review of this page: reversing raw waveform audio and then extracting onset features is not a fair symmetry test, because spectral flux measures positive frame-to-frame increase and reversal turns attacks into decay shapes. The control here reverses the already-extracted forward feature sequence, which is the defensible form. The identity conclusion also stands without any DTW: the release credits Ralph Humphrey and Tom Fowler at a listed 3:59, your stems are 238.14 s, and the s187082 clock describes a 277 s object.

4. The tab that does match

A song-name search surfaced a second Camarillo Brillo. Its tracks carry real personnel: Ralph Humphrey (Drums), Electric Bass (Tom Fowler), both correct for Over-Nite Sensation. The tab carries aiGenerated: false and runs 122 bars at 120 bpm stepping to 122 bpm at bar 10, which predicts 240.33 s against 238.14 s of audio.

Both tabs then went through one identical test. Each was mapped by its own tempo clock plus an affine rate, then given a per-four-bar drift correction bounded at a quarter second, which is half a beat and cannot slide the fit onto a neighbouring beat. Score is mean onset support across kick, hat and snare.

5. Verdict on the five instants

The more useful finding sits one level up. The same census run on s412176 returned zero three-surface instants across 1,769 note events. State the conclusion at the width the evidence supports: the five stacks are transcription defects in s187082 because they require impossible simultaneous stick assignments, and which component of each stack is the erroneous one still needs stem-level verification. A real ride plus snare could sit under an added tom, or a real snare plus tom under an added ride. Zero collisions in s412176 shows that its arrangement avoids the impossibility, and it does not by itself prove every component of those five stacks is absent from Humphrey's playing.

6. How good the matching tab actually is

Every written note in s412176 was mapped through the validated piecewise clock and asked its own stem for support. The chance rate is 0.05 by construction, so enrichment is the number that matters.

Four lanes clear the reference-class threshold at the lane level. Read that precisely: these numbers show significant aggregate correspondence between the written lane and its separated stem, and individual questionable events remain subject to direct stem verification. A lane statistic cannot certify a single note. Only 20.3 percent of written snares and between 10.6 and 13.9 percent of the other written events meet the support criterion, so the statistic establishes identity, guides navigation, and ranks suspicious passages. It does not stand in for hearing the attack at the written instant.

The cymbal lane sits at 1.3×, under the 1.5× floor, so all 193 written cymbal events were carried into a per-event pass, reported in section 7. The 182 in the table above is the subset that maps inside the recording; the other 11 fall past the end of the audio, which is the tab's 240.33 s prediction overrunning a 238.14 s file. Broadband cymbal texture is the known weak class for source separation, and this measurement behaves the way that literature predicts.

7. The per-event pass, and the null that was wrong

The lane statistic in section 6 was never going to settle a single note, so all 193 written cymbal events went through an event-level pass. Timestamps were tightened first: 546 kick and snare landmarks gave a sliding local correction that pulled the median residual from -36.0 ms to -1.2 ms, with 67 percent of landmarks inside 50 ms.

Two claims an external review knocked down, and the reruns

This page was sent out for adversarial review before it was trusted. Two objections were correct, both were runnable, and both were run.

The detector overcounts on every lane including kick and snare, and the cymbal ratio collapses to 1.3× at the strict setting exactly as kick collapses to 1.0×. The omission claim is withdrawn. What can be said is narrower: across every tested threshold, detected cymbal attacks outnumber written cymbal events, which is consistent with substantial omission, and detector oversegmentation prevents converting that count difference into confirmed missing notes.

The repair holds beat phase fixed. Each written note is now scored against the same eighth position in every other bar that carries no written note of that lane there. A downbeat crash is compared with other downbeats and never with its own bar's offbeats.

The principle from section 6 survives all of this intact and is worth keeping separate from the result. A lane-level enrichment measures association across many observations and never certifies an individual annotation. The name for the pattern is an aggregate versus event-level evaluation discrepancy, and MIR reports frame-level and event-level metrics separately for exactly this reason. The lesson is that an aggregate pass and an event-level pass are different tests, and here the event-level test needed two attempts before its null was sound.

8. The investigation queue, rebuilt on the corrected test

Because the lane now localises well, a low rank is a genuine outlier rather than a limit of the measurement. That turns a vague list of 182 unsettled events into a short list worth hearing.

All nine are crashes and none is a ride. Each was cut to audio, an isolated Cymbals excerpt and a full-kit excerpt, 0.45 s before the written instant and 0.75 s after. 177 events produced 354 files, named by bar and beat so a re-ranking never needs a re-cut. Queue at ~/Projects/_outputs/zappa-camarillo-brillo/cymbal_pass/INVESTIGATION_QUEUE.md.

9. Three rounds of review, and what survived all three

This page went out for adversarial review twice. Seven objections came back, every one was runnable, every one was run, and four of them changed a published claim. The table is the audit trail.

What the class-separated rerun shows, and it is the best result on this page

Separating the classes turns a vague result into a specific one. Not one of the 135 written rides falls below the midpoint of its own beat position, and 13 of the 46 written crashes do. The ride writing in this tab is excellent. The crash writing is where the doubt lives, and the earlier mixed-class analysis was hiding that behind an average.

The all-crash pattern is not a sampling artifact. With 46 crashes among 181 rankable events, drawing nine crashes as the nine lowest by chance has probability 2.3 × 10^-6.

What the nine themselves show, after two withdrawals

The routing objection, settled

One explanation could overturn every crash finding at once: if the separator writes real crash energy into the Hi-Hat output, then the surviving hi-hat evidence IS the crash. The full-mix attempt failed its control, so the question needed a different instrument. This one works, and it is simple. Take the crashes this tab writes that the stem strongly supports, treat them as ground truth, and ask where their energy actually landed.

Two of the nine are a different case. Bar 72 carries a real crash about 0.30 s after its written position and bar 43 about 0.45 s after, against an alignment residual whose median is -1.2 ms with 67 percent of landmarks inside 50 ms. Those two read as displacement rather than fabrication, and displacement is not vulnerable to the routing objection in the same way.

Evidence: round3_checks.json, round3_medians.json, independence_attempt.json, nine_suspects_spectrograms.png, and the per-event audio in cymbal_pass/audio/.

All thirteen low-rank crashes now read, and the outlier re-read by eye

The class-separated pass found 13 crashes below the midpoint of their own beat position. Nine were read first. The remaining four went through the same three measurements plus a spectrogram.

Bar 47 is the only one of the thirteen that lands between the two reference groups, since the threshold set by the routing test is at or below -5.8 dB for a supported crash and at or above +13.6 dB for the nine. Its +3.3 dB belongs to neither, so it stays unresolved rather than being pushed into the verdict that suits the pattern.

11. What s187082 actually is, settled from its own metadata

Every earlier section said s187082 does not describe this recording without being able to say what it does describe. One metadata read closes that, and no audio pull was needed.

The alternative row is the giveaway. Its clock starts the tab at -22.05 s against the vinyl mix, meaning 22 seconds of written music sit before that recording begins. A tab fitted to a longer live arrangement and then loosely attached to the studio version is exactly what that negative offset looks like.

It also retires a phrase used earlier on this page. Calling s187082 an AI tab that fails to describe the recording is accurate about the comparison and misleading about the tab. The accurate statement is that s187082 carries five physically impossible written moments and describes a different performance from the one these stems separate. The five stacks remain transcription defects in that file under the two-hands rule, and which note in each stack is wrong stays unknown.

12. Settled, open, next