# Zappa Drum Transcription Toolkit The skills on this Mac that read a drum part, plus everything already on disk about Frank Zappa's drum transcriptions, with **Keep It Greasey** pulled out in full. Built 2026-09-08. Song locked for this chat: Keep It Greasey, `s604777`. ## The skill stack, seven stages | Stage | Skill | What it does | |---|---|---| | separate | `/composite-stem-alignment` | One clean multitrack from several mic captures, one transform per mic. | | separate | `/multimic-stem-fix` | Repairs phase cancellation and bleed between kit lanes. | | read | `/perfect-stem-read` | The governing document. Stem event, instrument identity, score timing, lane last. | | read | `/drum-dynamics-from-audio` | Fixes velocity and articulation by measuring the recording. | | notate | `/stems-to-guitar-pro-drums` | Drum stems to a Guitar Pro file with flams and ghosts recovered. | | notate | `/audio-stems-to-midi` | Cheaper first pass, demucs plus crepe plus basic-pitch. | | repair | `/transcription-repair` | Human transcription as the model, stems as evidence testing single claims. | | repair | `/songsterr-drum-repair` | Restores a drum staff without damaging the rest of the tab. | | standard | `/notation-evidence-gate` | Settles a symbol against Weinberg, refuses a claim with no null model. | | grade | `/reference-class` | Decides whether a number earned the right to ship. | | publish | `/songsterr-upload` | Corrected Guitar Pro onto Songsterr, synced to the recording. | | publish | `/songsterr-tab-guide` | Data-driven tab-guide page for a band's albums. | Beside them: `/midi-drum-analysis`, `/five-stem-song-analyst`, `/stem-analyzer`, `/midi-to-tab`, `/album-tab-citation`, `/impossible-guitar-parts`, `/fraud-firewall`. ## The Zappa program, current state | Fact | Value | |---|---| | Zappa song ids in play | 34 | | Ghost flags stripped by the 2026-08-29/30 sweep | 3,842 across 16 live tabs | | Restore revisions held on moderation | 17 | | Printed charts on disk | 12 | | Published audit pages | 32 | Weinberg's 1994 *Guidelines for Drumset Notation*, **page 20, plate Ex. 11**, prescribes the parenthesised notehead the sweep read as an error, and its wording covers cymbals too. Read 2026-09-06 by rendering the image-only scan at 400 dpi, since `pdftotext` returns 12 bytes: `normanweinberg.com/uploads/8/1/6/4/81640608/940506pn_guildines_for_drumset.pdf`. ## Keep It Greasey Three Songsterr tabs. `s604777` is the one that matters: human, by Ben Dibden1, nine parts named by player, revision `r8852151`, **248 bars, 4,948 drum events, 629 ghost flags**, 8:18 against the album's 8:22. **Revision history.** Kevin Sheppard approved 2025-07-15, then a strip revision from Brandon's account on 2026-08-29 removed all 629 parenthesised noteheads and was marked **alternative**, so it never published. Ben Dibden1's revision `r8852151` is live, stamped 2026-09-02 local and 2026-09-03T01:08:51 UTC: section headers, rests cleared from empty bars, tracks renamed to the template format, bass tone set to finger. **The sweep hit this song and failed to land, so all 629 ghosts are intact.** **That strip's revision note is stale.** It describes what `r8768457` would have done, never the current state. A signed-in read on 2026-09-08 gives `s604777` 38 revisions, latest `r8852151`, `isOnModeration: false`, and the copy `s6862054` exactly one revision. Nothing newer exists. **Ghost layer, verified twice.** The live export carries 629 flags, **every one on MIDI 38, acoustic snare**, across 152 of 248 bars from bar 2 to bar 238, with beat dynamics ppp 12, pp 94, p 376, mp 33, mf 114. That reproduces the validation page's table by a different route. **Audio.** Fifteen stems at `/Users/Shared/202 Keep It Greasey-D minor-142bpm-442hz/`, kit already split six ways. The metronome stem is synthetic, a beat tracker rendered to a wav. **Printed sources, digits read from native-resolution crops at 4x and 6x, never a page overview.** DRUM! Groove Analysis reads **Verse 1 at 0:35, quarter = 134, 19/16** and **Verse 2 at 1:16, 21/16 into 9/16 into 4/4**. The Odd Meter Lesson page draws the ghosts in brackets. Marc Atkinson's Modern Drummer *Rock Charts* is the only complete-song chart, now re-rendered at 400 dpi. **Tempo.** Measured 136.7 BPM at the 3:18 passage against a printed 134. Speed shift is about 1 percent, so it does not explain a 6 percent gap. **Validation of 2026-09-07.** Snare and tom stems show two dynamic populations. The score agrees with itself: 629 flagged events sit 1.804 dynamic ladder steps below 794 unflagged, z = -30.58. The alignment never closed, so 0 of 248 bars carry a surviving two-direction result and **zero repairs were applied**. **Four open findings.** `KIG-M001` a GM 65 on the kick line at bar 231, `KIG-M002` 21 hand claps at the snare position, `KIG-M003` 114 ghost flags resolving to mf, `KIG-M004` 32 unnotated kick onsets. ## The copy tab is already staged `s6862054` "Keep It Greasey by Frank Zappa Brandon edit", artist string **Misc Covers**, carries one revision `r8958071` created via `Copy`, with no edit submitted. Its export matches `s604777` exactly: 248 bars, 9 tracks, 14,444 events, 629 ghosts, drum track 4,948 and 629. The public API returns nulls for it, so read it from the signed-in browser. ## KIG-M001 applied and published, 2026-09-08 Copy tab `s6862054` now sits at **`r8972975`**, createdVia GP, tracksCount 9, isOnModeration false. The original `s604777` was never touched. Bar 231 beat 7 went from `{"fret": 65, "string": 4}` to `{"fret": 35, "string": 4}`. String 4 on that staff carries 1,311 kicks and nothing else, so fret 65 there was an encoding slip. Gates: `diff_gpif_tracks.py` exit 0 with only Vinnie Colaiuta changed, `preflight_import.py` PASS, grid 1.30% inside tolerance, NAMEDROP-CLEAR on 9 of 9 CDATA names. The edit was a surgical text replacement on the raw GPIF, so all 149 CDATA blocks survived. Verified uncredentialed from CloudFront, http 200: fret 65 count 0 (was 1), kick 1,312 (was 1,311), ghosts 629 unchanged, 248 measures, 4,974 events, **all 9 track names intact**. ## The shared-element trap, measured 2026-09-08 The chain from bar to note on the KIG-M001 target was **1 bar to 1 voice to 1 beat to 1 note**, so rewriting the note in place could reach nothing else. The same file's worst sharing is one beat referenced by **739 voices** and one note by **144 beats**, so the trap is real here and simply did not touch this edit. Element ids do not survive a Songsterr round trip: the re-export moved Beats 1,527 to 1,556 and Notes 680 to 679 with no content change. Compare content tuples, never ids. **No count in this pass came from a fitted threshold.** Every figure is a direct field read, and the detector figures from 2026-09-06 and 09-07 are cited as those sessions' work. ## The eye check on KIG-M001 Bar 231 rendered from both files with alphaTab headless at scale 4.0 and diffed pixel by pixel. **BEFORE the notehead sits in the space below the bottom staff line; AFTER it sits on the bottom line**, with every other kick in the bar. The diff isolates a 42 px notehead-sized change at x 1122-1163 spanning y 513-565, straddling the bottom line at y 532. Everything else that differs is beam re-layout below the staff. The eye agrees with the fields, so nothing was overruled. Ghost and dynamic are **different elements**: ghost is `` on the Note, dynamic is `` on the Beat, and this file carries **zero** ``, so the Montana substitution failure has no counterpart here. The 09-07 page's 794 unflagged is snare-lane only (629 + 794 = 1,423 = the MIDI 38 count); all lanes gives 4,319. ## Id-histogram pre-flight on the two remaining edits | Plan | Note id | In plan | Total file-wide | Verdict | |---|---|---|---|---| | KIG-M002, 21 hand claps | 667 | 21 | **21** | **Exclusive**, in-place edit safe | | KIG-M003, 114 ghost flags at mf | 659 | 114 | **629** | **Shared**, 515 outside the plan, copy-on-write mandatory | Note 659 is the shared snare-ghost definition carrying all 629 ghosts. The first pass of this check compared instances against *beat-references*, 21 vs 5 and 114 vs 37, and wrongly called both exclusive. Counting instances on both sides flips M003 to shared. The worst-shared beat in the file, id 1020 at 739 voices, belongs to **Warren Cuccurullo**, a guitar track, so no drum edit can reach it. ## The bundle, collected 2026-09-08 `~/Projects/_outputs/zappa-keep-it-greasey-transcriptions/` holds all three Guitar Pro exports plus the pre-sweep baseline, every printed chart rendered and labelled by transcriber, and the nine live part JSON files. Two GPIF traps hit and fixed: `Notes` are shared definitions, and a ghost is a direct `` child rather than a Property. Walking the export correctly reproduces 4,948 events and 629 ghosts, matching the CloudFront read. The Pete Sweeney PDF holds **three** charts, not the two its manifest records. Palette: The Royal Tenenbaums (Wes Anderson).