Keep It Greasey: three verdicts

1. The reproduction check that came first

An independent walk of the published Guitar Pro file returns every published figure from a separate route, so the file on disk is the file that was measured.

2. KIG-M003, the premise that did not survive

The open item said 114 of the 629 ghost flags resolve to a dynamic of mf, that the two encodings disagree, and that nothing on hand decides which the transcriber meant. The first half of that is a reading error.

The 114 came from a resolved field, never an authored one

The Guitar Pro exporter resolves the sticky Songsterr value onto every beat, so reading the export hands each ghost a dynamic whether or not the transcriber wrote one there. Splitting the 114 by whether a mark was actually authored at that beat:

The control, and the null

Across the whole staff the transcriber authored 363 dynamics at ghosted notes. Every distribution below is a direct field count.

What the notation standard adds

Norman Weinberg's PAS drumset guidelines settle the symbol: a drumset ghost is a notehead change, never a point on a dynamic ladder. Guitar Pro naming the flag AntiAccent invites the ladder reading and the ladder reading is wrong. A ghost notehead inside an mf passage is well-formed notation, so the remaining 3 are not a defect either.

The transcriber's own words, in the file

Bar 22 beat 0 carries a text annotation:

The person who wrote the flags was tracking them bar by bar in this exact section.

3. KIG-M002, a change nobody could see

Twenty-one events on GM 39, Hand Clap, drawn at the snare staff position. The open item called the placement certain and the intent unrecoverable.

One, the figure is deliberate

The snare never thins inside those bars, running 4 to 8 events per bar throughout, so the claps are an added layer rather than a substitution. A stray GM number does not land on the same subdivision of eleven consecutive bars.

Two, the change is invisible to every reader

Songsterr agrees from its own side: all 21 claps and all 1,423 snares sit on string: 1.5, and its note schema has no notehead field. That is two field reads, so the question then went to the eye with a control.

The two source files differ by sha256 and the two render pages differ by sha256, and the control fires at the clap's own position in the bar, so the pipeline sees edits. Rewriting the clap as a snare produces a byte-identical PNG.

Published, GM 39

Control, GM 49 crash

Three, no independent printed witness notates a clap

Marc Atkinson's Rock Charts transcription is the only complete-song printed chart on hand. Its music key lists tom-toms, snare, bass drum, hi-hat by hand and by foot, ride, crash and additional toms, with no clap, cowbell or auxiliary lane at all.

2b. The velocity trap across four staves, and the limit on it

A peer session reproduced the sticky-velocity finding and sent a spread. Every figure was re-fetched here from uncredentialed CloudFront reads of the live heads, not accepted. These are direct counts of a stored field, not detector output, so no chance baseline applies to any row.

The limit, which carries more weight than the spread

Two reproduction details found alongside

Montana s35870 has 18 tracks and its drums sit at index 15, with further drum tracks at 16 and 17. A scan capped at 14 parts misses the drum staff entirely. The meta's popularTrackDrum field gives the index directly and was confirmed as 15 here.

The part path is https://dqsljvtekg760.cloudfront.net/{songId}/{revisionId}/{image}/{trackIndex}.json, where {image} is the meta's image field, a v0-3-2-... string. Passing the per-track hash instead returns 403 on all three tabs. Responses are gzipped.

3b. A third route reaches the same finding on M002

A separate session measured this song with a two-hands playability gate, which shares no method with the rendering test or the field reads above. Its claims were re-derived here rather than accepted.

Every value matches what the peer reported. Both versions of tr_score.py were then loaded here and run over all 4,948 events and 3,730 written instants:

Exactly the three beats claimed, at exactly the bars claimed, and they are the same collisions recorded above as (38, 39, 42) stacks. ./run_selftests.sh was also run here directly and returns ALL SUITES PASS, exit 0.

3c. One drum staff, no percussion staff, and what that removes

A peer session's track census reported this and it was re-derived here from the s604777 meta rather than accepted. Nine tracks, exactly one instrumentId 1024 staff at index 8 named Vinnie Colaiuta, and no staff named for tambourine, cowbell, shaker, conga, timbale or percussion anywhere. popularTrackDrum reads 8, matching.

The verdict on KIG-M002 is unchanged: no change to the tab.

The peer's own correction, recorded because it bears on other songs. That session had told the record that Peaches En Regalia s35889 has no separate percussion track. It has a staff named tambourine at index 10. Their census had searched for percussion among staves whose instrumentId was not 1024, so a percussion staff carrying 1024 was invisible to it. Their published instant counts were not affected, because those already measured every 1024 staff rather than the first, and they checked rather than assumed. Nothing on this page depends on it, since Keep It Greasey has a single 1024 staff.

3d. Staff-line numbers are PER-TAB, and this page had been treating them as stable

The peer reported that a line value means something only inside one tab. Built independently here from the drum staff of five tabs, uncredentialed CloudFront reads, taking each pitch's most common line:

The Carolina case, corrected

Carolina carries snare2 MIDI 40 at line 1.5 with 229 notes, so line 1.5 there genuinely is the snare line, with a real snare voice on it. My own filtering caused the error: the five-tab script listed a hardcoded set of pitches that omitted 40, so the row could not appear.

A recurring artifact the full table exposed

Listing every pitch also showed out-of-kit MIDI 65 sitting on the kick line in three of the five tabs: Alien Orifice 4 (2), Carolina 4 (3), Keep It Greasey s604777 4 (1). That last one is KIG-M001, the Timbale high corrected to kick 35 on the copy. The same artifact sits uncorrected in at least two other tabs, so it is a class rather than a one-off.

This explains the near-miss above

On s68253 both hatClosed 42 and tambourine 54 fall on line 1 across the tab, so a tab-level census shows them sharing. At each of the 118 instants the tambourine is still alone there, because hatClosed sounds only 5 times on the entire staff. Lane-level co-occupancy across a tab is not instant-level sharing.

The peer's resolution of the 172, recorded as theirs

Their per-instant run-length test resolved the wider set: of the 172 non-kit instants, 155 are a second player, 16 are ambiguous four-bar runs on Call Any Vegetable s68241, and exactly 1 is a likely real drummer flag, an isolated single tambourine bar in Brown Shoes s35869. True catalogue two-hands failure count 611 rather than the 822 a naive read would have shipped. Not re-derived here, per the one-song rule.

3e. My correction of the peer's cell was itself a units error

Their 133 was per-line and my 139 was per-pitch. Both correct, counting different things. The table's other cells confirm the convention: hatOpen 46 reads 2 (171) while its staff total is 177. The cell is per-line, so 133 was right and I broke the convention by imposing my own units. Reverted, and the note now states its cell format explicitly.

A pitch-to-line map is wrong inside a single tab

The unfiltered census exposed something stronger than the per-tab rule above. Six of s68253's eleven pitches occupy two lines each:

The same error shape hit both sessions in one exchange

My hardcoded pitch dict could not show midi 40. The peer's range(35,82) boundary could not show midi 65, because High Timbale sits inside GM percussion while sitting outside the standard drum kit, which is the whole point of that finding.

4. KIG-M004, still open, and by how much

Thirty-two kick onsets the score omits, listed by audio time only. Placing them needs an anchor, so one was fitted on kick and hi-hat only across bars 22 to 32, with snare and clap held out because the claps are the target and the snare shares their staff row and their stem.

The one thing worth carrying forward

The fitted 35.980 s sits within about a second of DRUM! Magazine's printed "Verse 1 at 0:35", a source that never entered the fit and shares no method with it. The score's own constant-tempo timeline says 38.217 s and drifts to 517.9 s against 501.8 s of real audio, so the notated tempo map is the thing that is off.

That agreement is recorded as one candidate landmark. The absolute bar identity gate requires twelve across two source families, six each, so no bar position is claimed from it. The obvious second family, sung-lyric onsets from the isolated vocals stem, is unavailable: the score's newLyrics field holds five entries and every text value is empty, so the lyrics would have to be entered by hand first.

4b. A second landmark family, found by not accepting the first dead end

Two families, fitted separately, sharing no stem and no notated part

They disagree by 156 ms. A 19/16 bar here runs about 2.09 s, so both land on the same bar. The 2026-09-07 gate failed its primary external-family requirement because lead and bass disagreed by 2.673 s.

The integer-shift contest

Zero shift wins separately in both families, by +33 and +24 matches over the nearest competitor. That is the condition the earlier pass failed, where kick, snare, hat and tom all leaned to plus one bar.

The gate asks for a median under 75 ms and a p90 under 150 ms. Both families clear both, inside the fitted span.

What still blocks M004, precisely

The anchor is local and it does not extrapolate. Carried across the whole recording it collapses to chance.

Net movement: bar identity at bar 22 now has two-family corroboration and wins its shift contest, where it previously had neither. Placing the 32 events needs a piecewise map with landmarks bracketing each one.

4c. The clap timbre test was run, and its own control killed the interesting part

Pre-registration v2, sha256 383fe83e...25f0b7, written with no timbre yet measured. v1 stays failed and none of its thresholds were moved. v2 rests on the two-family anchor that did not exist when v1 was written, and declares the same timbre thresholds v1 declared.

The declared test returned UNEVALUABLE

Entry rule fixed in advance: an instant enters only if a detected snare-stem onset lies within 60 ms of it.

Zero of eleven entered, so neither the centroid nor the low-band comparison could run.

The post-hoc observation, and the control that withdrew it

Not pre-registered, and reported here because it looked like a finding. The 11 clap instants scored 0 in the snare stem while the same-bar notated snares scored 33 of 58 = 56.9%, giving a binomial p = 9.5e-05. That baseline was circular, comparing beats with no notated snare against beats that have one.

For completeness, 9 of the 11 clap instants do carry an onset in at least one of the seven stems, so the anchor is not wrong at those instants. Twelve of the 21 claps stack with a notated kick, which accounts for the hits in the kick, cymbal and other lanes.

4d. The piecewise map was built, and it failed its own external check

Pre-registration sha256 2e328671...3e80f10, hashed in advance of any window fit. M004 was waiting on this step, so it was run rather than queued.

Sixty windows, 8 score bars wide stepping 4, each fitted independently on kick and hat only, offset searched across the whole 501.84 s, scale 0.93 to 1.07, one-to-one at 50 ms. Fifteen windows passed the four declared criteria and placed 64 surplus kick onsets across 48.3% of the recording.

Then the control ran

The same two-family logic that made the bar-22 anchor credible, applied per window at that window's own offset and scale, tolerance 60 ms: does the isolated vocals stem corroborate it?

One of three testable windows corroborates. The other twelve sit in the instrumental solo and carry too few vocal attacks to test, so they are UNEVALUABLE rather than passing.

The lesson, which is the useful output

An 8-bar window carries too little information to lock uniquely across 501.84 s of a dense, repetitive drum part. Spurious locks cleared the declared 2.0x rotation-null threshold, so 2.0x was too loose at this window size. Raising it alone is not the fix, since refitting a declared threshold on the same data is the error these gates exist to stop.

1. A window must be corroborated by a second family to place anything. Only 3 of 15 were testable, so the vocal part cannot reach the solo section and a third family is needed there.
2. Monotonicity must be enforced against passing neighbours, and the code must assert a window's offset lies in a plausible band for its bar range. Bar 210 at 42 s should have been rejected ahead of any statistic.
3. The search should be local. A free full-song search per window invites the spurious lock that produced +46 s and -65 s.

4e. The third family exists, and finding it moved the blocker

The previous section named the next step as find a corroboration family reaching the instrumental solo, since the vocal part stops at bar 101. That search was run rather than queued.

The "no family reaches the solo" blocker is dissolved. Two candidates cover it, and a third covers the solo passage itself.

The bass cross-check, and why it is not carried as a refutation

The bass family made all 15 previously-passing windows testable, where the vocal family could test only 3. Every one of the 15 disagreed with its window's drum offset, by between 3.985 s and 405.500 s.

All eight clear the 1.5x floor, and all eight clear it narrowly, against a surface median near 51%. Gate N4's own warning covers this: above a 35% null a small excess goes significant while meaning little. Compare the whole-part vocal fit at 7.65x and the bars 22-32 drum fit at 2.01x. A second contributor: the score notates 2,424 bass attacks while the bass stem yields 1,825 onsets, so 599 can never match and a perfect map still caps near 75%.

4f. The window-size fix is superseded by a named method

What was built here was the wrong shape. Sixty windows, each fitted by an independent free search across the whole 501.84 s, with nothing tying one window's answer to its neighbour except a monotonicity check applied afterwards. That is how a window covering bars 210-217 came back at 41.980 s. Widening the windows would have reduced that failure without removing its cause.

The named methods

None of the four was run in this pass, so for every one of them the offset, scale, tolerance and match rate are all not applicable here.

• Multiscale DTW (MsDTW), Müller, Mattes and Kurth, ISMIR 2006. Coarse-to-fine: a low-resolution result is projected onto the next level and refined inside a bounded region, which supplies exactly the neighbour bound this run lacked.
• Path-constrained partial synchronization, Müller and Appelt, ICASSP 2008. Built for the case where only parts of two sequences correspond, which is this case, since the drum staff runs 248 bars while any one family covers a subset.
• Subsequence DTW and common subsequence matching, FMP C7S3. Matches a short query against any subsequence of a reference, the right primitive for one region without a whole-song map.
• Music synchronization with chroma features, FMP C3. The chapter the whole task belongs to.

Two setup facts, corrected after trying

The FMP source text is catalogued in this project at /Volumes/Black/complete/Springer/ and that volume is not mounted on this machine right now, so the chapter was not read here. The method names, authors and venues come from the search.

Neither gap blocks the next attempt. MsDTW needs only numpy and scipy, and the FMP notebooks are readable on the web at the links above.

6. The reference class this result sits in

A peer session tested the completeness of the 38-id scope list the Zappa program works from. The two headline figures were re-derived here by paging /api/songs directly, not accepted.

What this does to the Keep It Greasey result

What it does change

It changes what a clean result would mean if anyone generalised from the program. The 38 ids are not a random draw from 262. A tab enters that list because someone had a reason to look at it, so the set is selected, and a low defect count across selected tabs licenses nothing about the other 230.

The 822 must not be read undecomposed, because three different things were being summed:

Corrected kit-only ranking: Flakes s35897 270, Tryin' to Grow a Chin s68245 78, Brown Shoes Don't Make It s35869 54, Filthy Habits s68237 24, King Kong Itself s35863 21, Let's Move To Cleveland s68238 18, Sinister Footwear II s21874 16, Mother People s60347 16, Carolina s68248 15, Dog Breath Variations s68247 9.

Two structural claims I did re-derive, because this page republishes them

The pitched-percussion artifact is the same class as KIG-M001

A staff named "Percussion Organ" carrying instrumentId 1024 holds melodic pitches while wearing the drum instrument id, and the sweep counted its events as drum-kit failures until the peer caught it. That is 40 of the 822.

What the 172 does to KIG-M002

The clap question looked local when it was 21 events in one song. The finished sweep puts 172 of the 822 instants on the tambourine, clap and cowbell question, and The Idiot Bastard Son s68253 carries 118 instants of which zero are kit-only, with 112 of the 118 being hatOpen plus snare plus tambourine.

Why Keep It Greasey's case works, and the 172 do not

At KIG bars 42-44 the GM 39 clap sits on string 1.5, the same line as the snare, with a snare sounding at the same instant. Two surfaces either way, so removing the clap changes nothing. Measured here on the largest affected tab, The Idiot Bastard Son s68253, revision 704236, drum staff index 3, 817 events, uncredentialed CloudFront read:

A sample instant reads (54, line 1), (46, line 2), (38, line 4), which is tambourine, open hat and snare on three separate lines.

The lane position varies per tab, at 0, 1, 1.5 and 5 across the five affected songs, which is itself evidence the transcribers gave the tambourine a deliberate separate lane rather than folding it onto a kit surface.

Nothing about Keep It Greasey changes. M002 stays closed at no change, and the 21 GM 39 events stay exactly as Ben Dibden1 wrote them.

4g. Step 5 started, and the method reproduces the known result

The method was named earlier and the work deferred at that point. That deferral has ended: libfmp now imports, so the subsequence-DTW primitive from FMP C7S3 was run here on the one window whose answer two independent families already agree on. The criterion was fixed in the script ahead of reading the result: the recovered bar-22 downbeat must land within one 19/16 bar, 2.095 s, of 35.980 s.

Setup. Reference is the pooled kick and hat onset envelope across the whole 501.84 s at a 50 ms hop, 10,037 frames. Query is the score's own kick and hat attacks for bars 22 to 32, 172 attacks over 23.4 s, 470 frames. Parameters for every row below: offset is what the run recovers rather than an input, scale is not a parameter of this method since the path is free to warp, tolerance is the 50 ms frame hop, and the match rate is replaced by a mean accumulated cost of 0.0183 over a 511-step path.

Path spans reference frames 726 to 1182, 36.300 s to 59.100 s, a 22.8 s span against the 23.05 s the score's eleven 19/16 bars occupy at 136 BPM.

Why this carries more than a fourth number

The two matched filters share a model, one offset and one scale. Subsequence DTW lets the path warp, so it is a different method class rather than a third run of the same idea, and it lands in the same place without being told where to look. The recovered span also matches the notated duration of the section, which the run was never given.

7. A provenance error of mine, and KIG-M001 verified properly

What it does to the figures here: nothing, and why

Every other column is identical between the two files, because KIG-M001 touched one note in bar 231. Snare 1,423, clap 21, ghosts 629, beats 3,756 and velocity beats 989 all hold on both, so the clap and dynamics tables are correct as published.

What it does to provenance: one row was mislabelled and is fixed

The four-stave velocity table credited 989 of 3,756 to 6862054. The measurement came from 604777. Corrected in place; the number never changed, the label was wrong.

The MIDI 65 class is six tabs, not three

My five-tab census found Alien Orifice 2, Carolina 3 and Keep It Greasey 1, all three exact against the peer's full 262-tab rescan. The two it could not reach are the large ones: Seal Call s604408 at 33 and The Deathless Horsie s35885 at 28, and on Seal Call the timbale outnumbers the kick on their shared line, 33 against 25. A five-tab sample sized the class at half its real membership, which is the reference-class point applied to my own finding. Those counts are the peer's and are not re-derived here.

8. The KIG-M001 repair is provably minimal, and that yields an acceptance test

Both sessions censused the two files independently and agree on every cell.

The acceptance test for any repair of this class

All six rows are arithmetically consistent, and three of the base counts were confirmed by my independent five-tab census.

Why a full census catches what a targeted check cannot

A check aimed at the edited note cannot see collateral damage. A peer session earlier lost 179 note instances across 79 bars by editing a shared GPIF Beat in place, and the edit looked correct at the target. A full census would have shown a count moving that nobody intended.

9. An attribution I got wrong, and the standard's real provenance

The 179 is grounded, checked here

reference_gpif_note_beat_dedup_copy_on_write documents it: removing 15 refuted notes from CODA-s6857183-Brandon-edit.gp by editing shared beats in place deleted 179 note instances across 79 bars that were never targeted, because <Beat> id 899 is referenced by 176 voices. The damaged file was quarantined rather than deleted. The standard's strongest claim rests on a measured case, not an anecdote.

The correction makes the standard stronger

A pre-write check the note was missing

The census runs after an edit. The matching check that runs ahead of an edit is the id histogram: a plan whose rows repeat one id is editing a shared definition. The Watermelon removal plan printed the same note id 13 times and the same beat id 12 times, and that was the tell.

The peer independently verified the note against their own measurements: the seven-row KIG table matches cell for cell, the six-row acceptance table holds arithmetically, both wikilinks resolve, and the MEMORY.md index line is correct.

10. Read the source, do not grep it, and a citation of mine that proves it

The peer reported they had verified the grounding note and found it sound. They had grepped it, not read it, and the pattern 179|79 bars|shared|in place matched none of the terms in its most actionable sentence. They caught that themselves and said so.

Both came from reciting what I remembered reading rather than re-opening the file. Corrected in the handoff, the only place they appeared.

Four instances of one failure, in a single exchange

The rule that generalises it

Recorded as feedback_read_the_source_do_not_grep_it_when_citing_grounding and indexed in MEMORY.md, since it applies well beyond this project.

The line that was nearly lost

Histogram ahead of the write, census after it. For the five outstanding MIDI 65 repairs both apply, and the histogram counts most on Seal Call s604408 at 33 notes and The Deathless Horsie s35885 at 28, the two large enough that a shared-beat reference is plausible.

11. A provenance slip inside the note that argues for provenance

The peer read the new rule note end to end rather than grepping it, which is the only honest way to check a note whose rule is "read it, do not grep it". Their verification, re-checked here:

What reading it caught that a grep would have passed

The fourth table row read: "reciting a line number from memory | that the line was 67, not 65, and the file 73 lines, not 74".

Why my instance is the worst of the four

A related shape: a check positioned where it cannot act

12. A third recitation slip, inside the message documenting recitation slips

The peer read the note again end to end and flagged one figure. They are right: I wrote that the note was 64 lines and it was 63.

The operational fix, which "read the source" does not cover

Reading a document is the rule for citing it. For a number about a document the rule has to be stricter, because a number can be recited long after the file was read and nothing in the text signals that it has drifted.

The peer's catalogue sweep, final, recorded as theirs

262 of 262 Zappa drum tabs measured, 611 real two-hands failures after decomposition, and 172 of 172 non-kit instants resolved to 155 second-player, 16 ambiguous, 1 real. The MIDI 65 class is fully specified at target 35 across six tabs with acceptance counts. Not re-derived here, per the one-song rule.

13. Two figures here are superseded, and a plan I queued was wrong

Reading the standing queue, which I never did

Measured: 216 open Zappa and Songsterr items, 141 predating today. I added items all session without once reading what was already there. Two consequences, both mine.

One: a plan I queued was wrong on three of six tabs

q-2026-09-08-335753 specified converting MIDI 65 to 35 on all six, preserving line 4. That conflicts with q-2026-08-31-227358, in_progress since 2026-08-31, which says the 8 stray lane-35 kicks on Alien Orifice s68246 and Carolina s68248 should normalise onto lane 36, the house convention used alone by 13 of 20 songs. My plan adds notes to the lane that item wants emptied.

Two: this page published a superseded count

It carried 610 kit-only and 611 catalogue failures as current. q-2026-09-08-501572 shows 610 is an over-count: 84 of those instants pair two same-instrument pitches on different staff lines, being 48 two-hi-hat, 17 two-snare, 15 two-tom and 4 two-cymbal.

The rule now covers more than documents

14. A count is a timestamp, and my 329 census carries the same filter defect

The queue disagreement resolves, and it is entirely filter definition

Two sessions counted the same shift_queue.jsonl and reported 216 open against 141. Measured again here at 2026-09-09T03:34:47Z, with the pattern pasted rather than named:

The rule this yields, theirs, stronger than mine

My 329 catalogue count has the same defect

This page publishes 329 Zappa tabs and 262 with a drum staff. Both came from paging /api/songs?pattern=Frank+Zappa and filtering client-side on artistId == 5912. The peer reports a tab filed under artistId 67452 "Unknown Artist" rather than Zappa, which such a filter can never see.

I could not verify that tab uncredentialed. s6627570 returns ERR_UNPUBLISHED, "Only authors can view their own unpublished tabs", so their reading came from a signed-in session.

Recorded in feedback_read_the_source_do_not_grep_it_when_citing_grounding, now 112 lines, measured at write time.

15. The 329 undercount has a measured size, and I repeated the filter error finding it

The peer showed the artistId ceiling is reproducible without credentials, so it was re-derived here rather than left queued. Measured 2026-09-09T03:42:05Z, rule stated rather than named: artist name matches zappa|mothers\s+of\s+invention, case-insensitive, over 1,705 rows paged from /api/songs for Frank+Zappa, Zappa and Mothers+of+Invention.

The residual is classification, not measurement. Of the 77, some are plainly other acts: Dweezil Zappa at 16, Zappa plays Zappa, Ozzy Osbourne & Dweezil Zappa, and two surname collisions, Zappacosta and Alfie Zappacosta. Others are plainly Frank under a variant spelling: frank zappa, Frank zappa, Frank ZappA, Frank Zappa., Frank Vincent Zappa, Frank Zappa 2. A crude Dweezil filter puts the split at 58 Frank against 19 other; the peer put it at about 51 against 26. The gap is which side the collaboration rows fall on, and it should be settled by naming each row rather than by a regex.

I repeated the filter error while measuring the filter error

What stays queued

q-2026-09-08-2f523c had two halves. The reproducible half is done and is above. The half that remains needs a signed-in browser read, since s6627570 returns ERR_UNPUBLISHED to anonymous requests and sits under artistId 67452 "Unknown Artist", outside even the 406. Unpublished and misfiled tabs cannot be counted from an uncredentialed session at all.

16. The song-wide pass ran. The map holds at bar level and fails at note level.

Pre-registered and hashed with nothing yet computed, sha256 47adf4eb...533b90a. This is the step KIG-M004 had been waiting on. Parameters, stated once for every row: offset recovered rather than supplied, scale is not a parameter since the path warps freely, tolerance 60 ms at note level, one-to-one match rate 510/1,311 = 38.9% against a 25.1% rotation-null baseline.

Coarse pass at a 200 ms hop, 2,595 score frames against 2,510 audio frames, 6.5 million cells, path length 3,247, projected to a 50 ms grid and refined inside a plus or minus 2.0 s band, the constraint region the piecewise run lacked.

Then the placement step returned an impossible number

The real measurement, one-to-one on the mir_eval convention

Rotation null at 60 ms, 300 draws: mean 25.1%. Observed 38.9%, gain +13.8 points, z = 9.13. verdict_gate.py returns UNDETERMINED, at 1.55x against a 2.0x floor.

What this pass did settle

17. The subband representation was built and run. It helps, and it is not enough.

Log-compressed spectral novelty per subband, FMP 6.1.2, replacing the pooled envelope: STFT 2048 hop 256, log(1 + gamma * |X|), positive first differences summed inside a band, local-average subtraction, half-wave rectify. Kick band 30-180 Hz at gamma 100, hat band 4-16 kHz at gamma 10. Every row below is a full coarse-to-fine map at offset recovered, no scale parameter, 60 ms tolerance.

Gain +17.1 points, z = 11.35. verdict_gate.py: UNDETERMINED, under its 2.0x floor.

18. I found a bug in my own map, fixed it, and the fix made the result worse

msdtw_v3.py carries no repair and claims no landmark evidence. Its parameters: 20 ms grid, plus or minus 1.0 s band, 2,646,390 cells, backtracking enabled, offset recovered, no scale parameter, 60 ms tolerance.

Why, and the finding in it

Where KIG-M004 stands, with both levers measured

19. The "chases noise" diagnosis put to the eye, and the eye agrees

This is the first eye check run on any map in this lineage, and it was run because a gate asked whether the detector had been put to the eye. It had not.

20. The path-shape lever, tried two ways, separates two different problems

Two constraints added to the fine banded DTW: a step-size restriction capping consecutive same-direction runs, and a wider envelope bump at fine resolution, on the hypothesis that a 20 ms grid leaves the onset envelope too sparse for the cost to discriminate.

The two problems come apart

What that points at, and it is not the map

A 200 ms grid beating a 20 ms grid on note matching is not a resolution story. The remaining hypothesis is that the cap is in the score rather than the alignment: the notation is quantised and the performance is not, so a 60 ms window against quantised positions may simply be too tight for the 61 percent that miss. That would make the miss rate a property of the transcription, not of the map.

21. The quantisation hypothesis is refuted, and the misses decompose

Section 20 raised one hypothesis for the capped match rate: quantised notation against an unquantised performance, which would show as misses clustering off the beat. Tested, and it is wrong.

The misses decompose: mostly detection, partly map

One localised map failure, and it has an address

20 bars of 197 match nothing at all, and the largest are contiguous: bars 204, 205, 206, 208 and 218, each with 6 to 8 notated kicks and none matched. Twelve bars match everything.

What this leaves

22. The missed positions are near-silent in the isolated kick stem

For every notated kick, the isolated kick stem's peak level was measured in a window spanning 20 ms ahead of its mapped position to 60 ms after, and the matched and missed populations compared.

The stem's median sample level is -67.4 dB, so a missed position at -57.59 dB sits about 10 dB above the noise floor of the file.

What this settles

What it does NOT settle, and the gate that says so

23. The map was fitted with the thing it was measuring

Section 22 left two readings open and named the full mix as the measurement that would separate them. A kick the separator failed to route into the kick stem would still leave low-frequency energy in the mix, so the mix can tell a silent performance apart from a silent stem.

The bandpass had to be repaired first

The full mix, under the same map section 22 used

Two hundred circular rotations per cell, seed 20260909. Read on its own this says the 752 missed positions are quieter than a random instant in the song, in the full mix as well as in the stem, which would have settled the question.

The confound that reading carries

The independent map

The same machinery built from hats alone, 1,156 notated hat events against 942 detected hat onsets, gives the 1,311 notated kicks a placement that never saw a kick onset. Under it, 450 of the 1,311 notated kicks match a detected kick onset within 60 ms.

What this corrects in section 22

Read the direction of that correction carefully. Section 22 reported the missed positions as quieter than chance, which reads as evidence that nothing was played there. The independent map returns them to chance, which is a null result and carries no verdict about the record either way.

The number that closes KIG-M004

Three hundred rotations of the mapped notated positions against the fixed detector output, one-to-one matching on the mir_eval.onset convention, 60 ms tolerance. Both runs of verdict_gate.py --counts return UNDETERMINED against its 2.0x counts-mode floor, at p = 1.55e-41 and p = 2.19e-13.

Exactly which sections inherit the limit, and which do not

The claim that downstream figures inherit this limit was checked against this page rather than asserted. Six sections above carry a ratio measured through the fitted map, sections 16 through 21, and each now opens with a banner naming this one. Section 22 is corrected in the table above.

Standing result

24. The census filter, measured against the staff it was filtering

Section 14 recorded a hardcoded pitch list as a defect and left the reusable guard unbuilt. The guard is built now, at ~/.claude/skills/_shared/census_guard.py, and its self-test replays the real failure against the real part file rather than a synthetic fixture.

What the literal cost, measured

The Keep It Greasey drum staff holds 4,948 notes across 17 distinct pitches. The literal {35, 36, 38, 42, 46, 49, 51, 54} declares eight keys, six of which are present.

Hidden pitches with their note counts: 39(21), 41(14), 43(112), 44(195), 45(85), 47(170), 48(109), 50(3), 53(282), 55(1), 65(1).

The exhaustive enumeration

Every note's own string field read, never a pitch-to-line lookup.

The audit reaches section 23 from the source side

Run across the 18 archived scripts in the working folder, --audit flags 12 literal key sets across 12 files of the 18 scanned, and 11 of the 12 are the same expression:

25. The acceptance test, built and controlled in both directions

Section 8 established that a repair is provably minimal only by full census diff, and left the reusable test unbuilt. The test is built now at ~/.claude/skills/_shared/repair_census_diff.py, composed on top of the census guard from section 24 so its enumeration is exhaustive by construction.

It censuses 8 scalar dimensions (measures, beats, notes, beats carrying velocity, rests, ghosts, dead notes, ties) and 5 keyed dimensions (fret, string, fret_string, velocity, duration), then reports every cell that moved. A repair declares its targets as --expect fret:35:+1, and anything else that moved is collateral.

The real repair, s604777 r8852151 against s6862054 r8972975

Its own vacuous-pass hole, found by testing it against nothing

A second hole in the same pass: a malformed --expect was only reported after the full diff had printed, so a typo in the declaration looked like a completed run. Expectations are now parsed ahead of any fetch or census.

Six exit codes, all measured in one command

Line 1.5 holding 1,444 notes is itself a check: snare 1,423 plus clap 21 equals 1,444, so the two lanes that share that line account for it exactly.

One of those six exit codes was published without being measured

Third instance of the same class on this page, after the audit file count in section 26 and the count in section 24. A number written from intent reads exactly like a number written from measurement, and only the measurement tells them apart.

26. A fourth recitation slip, in the section about filters, caught by measuring

Section 24 first published the audit result as "12 literal key sets in 11 files, every one of them r["fret"] in (35,42,44,46)." Both halves were wrong, and both were written by reading the output rather than measuring it.

The twelfth hit is score_timeline.py:58:

27. The energy result is withdrawn. The null never reproduced the selection step.

An external check on this page named the defect ahead of any measurement here: "If each rotation moves the positions while retaining the original matched and missed labels, the surrogate does not reproduce that selection step, and the observed missed group can look unusually quiet partly because membership was conditioned on failure to encounter a detected onset." The measurement below was run to test that claim, and it confirms it.

What was wrong

The corrected null

Each surrogate now rotates the candidate positions, re-runs the one-to-one 60 ms matching against the fixed detector, rebuilds matched and missed membership, and only then computes the energy statistic. 200 surrogates, seed 20260909, hat-only map.

Surrogate match rate 0.254 mean, sd 0.014, against an observed 0.343, which independently reproduces the 25.3% baseline the alignment ratio uses.

What collapses

Three of the four cells fall to chance. The fourth, the missed group in the kick stem at z +2.09 with 2.0 percent of surrogates above it, is weak tail evidence at 200 draws across four related cells, and it is technically an above-null result rather than a chance-centred one. Its sign runs against the absence argument rather than for it: those positions carry somewhat MORE kick-band energy than the reproduced selection null predicts, the opposite direction from the near-silence section 22 published.

What survives

The rule this replaces

28. The non-drum map. The alignment does not survive full independence.

This runs the strongest of three controls named by an external check on section 23: "A sufficiently accurate map derived from vocals, bass, guitar, or other non-drum landmarks would provide stronger alignment independence because drum co-occurrence would leave the fitting stage entirely."

Notated vocal attacks were built from the vocal part s604777_r8852151_part0.json through the same constant-tempo walk that produced the drum timeline, using the drum part's tempo automation since both parts share one score. 437 notated vocal attacks against 521 detected vocal onsets on the isolated lead stem, and the 1,311 notated kicks mapped through that path. No drum event of any kind enters this fit.

This is not a power failure

At n = 360 with a null of 23.9% and a rotation sd of 3.7 points, the test detects 1.31x or greater at z = 2.

The limit on this result

Standing result

29. The vocal map is invalid. Both of its figures are withdrawn.

The external check named this risk within minutes of section 28 being published: "the sequence alone cannot distinguish that explanation from a degraded third alignment whose input sequences cover substantially different spans." It is a degraded alignment.

The one-line validity check

A score-to-audio map must have a slope near 1.0, because score time already carries the tempo map. Anything else means the path is absorbing time it should not.

Why the vocal part cannot carry a song-wide map

Total measure time is identical to four decimal places, 498.1647 s for both parts, so the walk is sound. The vocal part simply stops. It covers the first 39 percent of the score and holds zero events after that, while the detected vocal onsets run to 285.6 s. No DTW variant can reconcile a 195.8 s query with a 501.8 s reference without distortion, and the slope numbers are that distortion measured.

What is withdrawn

What survives, and it is the useful part

One word narrowed, and it names a live route

The slope check is a REJECTION guard, not an acceptance criterion

A separate discrepancy, recorded rather than resolved

30. The bass map. A valid non-drum alignment, and the score clock explained.

This executes the route section 29 left open, and it passes its slope check.

First, the 19.62 s score-clock discrepancy is explained

The original score_timeline.py advances the bar clock with t = max(t, tt), the longest voice's actual beat sum. A second walk advances by the nominal signature. On this tab those diverge, and the reason is measurable.

The bass map

Arthur Barrow's part, s604777_r8852151_part4.json, on the nominal clock, against librosa spectral onsets on the isolated bass stem. No drum event enters this fit.

The result

300 rotations, one-to-one 60 ms matching, membership regenerated inside every draw.

How 1.17x should be read

The bass map removes drum events from the fit, which kills the circularity behind 1.67x and reduces the rhythmic dependence left in 1.36x. It cannot remove the fact that Barrow and Colaiuta are playing the same composition in the same meter, with bass attacks and kick attacks often organised around related beat positions. A whole-song circular rotation deliberately breaks that phase relationship, so a correctly registered score can outperform rotations even where the particular written kick notes carry little extra information.

On the statistics: the empirical p of 0.0166 and the z of 1.82 do not conflict. A rotation distribution need not follow a normal tail, and 300 draws make an empirical z coarse. A surrogate test is interpretable only through the structure its null preserves and destroys, and this one destroys metrical phase along with everything else.

Standing result

5. Files and reproduction

Onsets were detected on isolated stems, never the mix, at 44.1 kHz over 501.84 s, hop 256, log-compressed spectral novelty per FMP 6.1.2. The kick count reproduces the earlier surplus finding independently, 1,332 detected against 1,312 notated.

~/Projects/_outputs/zappa-keep-it-greasey-transcriptions/04-clap-test/ holds the pre-registration, its sha256, the results, every script and every evidence image.

Live tab: Keep It Greasey, Brandon edit, s6862054 · source revision r8972975, unchanged by this pass.