The Zappa ghost note debacle

A sweep across Frank Zappa drum tabs on Songsterr read genuine ghost notes as a notation defect and stripped them. This is the verified record of what was changed, what moderators caught, and what is still live.

Audited 2026-08-30 · moderation, track names and revision state re-read uncredentialed 2026-09-07 against api/meta/<songId> and api/meta/<songId>/revisions

1. The moderation wave2. The restore emptied the credits3. Repaired behind us4. The seventeenth tab5. Defects the restore introduced6. The upload path changed7. The work that is not ghosts8. The seventh instance9. What happened10. The original ledger11. Root cause12. Which lane13. Charts14. Accents

1. The moderation wave, 2026-09-07

Sixteen restores went up on 2026-09-05 and sat unread for two days. Moderators cleared them in a single afternoon. Every row below was read from api/meta/<songId>/revisions uncredentialed on 2026-09-07.

2. The restore emptied the credits

The repair for the ghost notes introduced a different loss on the same tabs. Eight restores published with their track names blank, taking authored musician credits off public pages.

The mechanism

GPIF stores a track name inside a <![CDATA[...]]> block. Python's ElementTree drops the CDATA wrapper when it writes the file back out. Songsterr's converter then finds no name and falls back to a generic label, so any restore serialized through ElementTree arrives with its names gone. Files built by direct download or by lxml with XMLParser(strip_cdata=False) keep them.

Four measurements across two songs isolate the variable:

The same upload path served all four, so the build tool is the variable and the platform is not.

Which restores dropped names, measured revision by revision

The three clean rows were built from direct downloads rather than through ElementTree, which is the same split the CDATA measurement predicts.

What is missing from public tabs right now

The fix is proven on two songs

Export the live revision, rebuild it with lxml using etree.CDATA under XMLParser(strip_cdata=False), verify zero note, dynamic and event changes, then submit. r8932311 did exactly that on Zomby Woof (AI) and restored 5 of 5 names, which is two better than the 3 the tab carried before the sweep touched it. RESTORED-s35881-TEXTSAFE-v2.gp carries 137 CDATA blocks and is the ship candidate for Watermelon.

3. Someone is repairing behind us

Within about an hour of each restore clearing, an Editor-built revision appeared on five of the songs. Three of them put names back.

Those repairs were made by hand in the Songsterr editor by people who did not cause the loss. Two tabs are still waiting for someone to notice.

4. Trouble Every Day, the seventeenth tab

This page recorded s412170 as superseded and safe. It was neither.

Our strip r8766285 removed 9 ghosts from the Chad Wakerman Percussion staff. Kirill's r8815000, which went live on 2026-09-01, put back 3. Six were still missing, all fret 49 crash at bar 67 beats 4 to 9, and all six notes still existed, so only the ghost flag was needed. Six wrong beat velocities were also still live: bar 1 beats 4, 5 and 6, bar 67 beats 4 and 6, and bar 69 beat 10. Twelve corrections in total.

RESTORED-s412170.gp was built and verified against the live base rather than the author, so Kirill's post-strip edits survive: 6 ghosts flipped false to true, 0 ghosts lost, 0 other note changes anywhere, 0 changes on other tracks, and the effective dynamic matches the author on all 807 slots. It went up as r8932027 on 2026-09-06 and is on moderation.

5. Two defects the restore itself introduced

The ride definitions were merged, so a naive flag would ghost everything

GPIF stores note properties in shared definitions. On Watermelon the post-sweep r8768414 merged ride definition 431, the ghosted one, and definition 432, the plain one, into a single id 418 covering all 1,724 occurrences. Writing <AntiAccent> onto 418 would have ghosted all 1,724, including the 408 that were never ghosted. The repair has to split 418 by occurrence position against the pre-sweep revision, copy-on-write, and log definition-level and occurrence-level changes separately.

A chord symbol has two halves and only one came back

GPIF stores a chord in two places, an <Item> defining the diagram and a <Chord> reference on the beat that displays it. The paren sweep dropped both halves of E7Msus2 at bar 3. The first restore put the Item back alone, so the definition sat orphaned and no bar printed the symbol.

Three checks passed on that broken file:

• 'E7Msus2' in text, which an orphan satisfies
• the printed list of chord Items, which never says whether a beat references one
• the lane-aware set difference, which keys on (bar, voice, position, midi) and cannot see a symbol carrying no notes

Beat id 2 is shared by seven voices covering track 0 bars 3 to 9, so the repair clones it copy-on-write and repoints the bar 3 voice alone. This is the same shape as the ghost loss itself: three checks agreed, and none of them measured the thing that was gone.

6. The upload path changed

For Watermelon that makes the action concrete: open s35881, create a copy, rename it to Watermelon In Easter Hay Brandon edit, import RESTORED-s35881-TEXTSAFE-v2.gp, and submit on the copy. The original tab and r8908034 stay untouched.

The preflight NAMEDROP rule is too absolute

The rule in preflight_import.py states that the uploaded file cannot prevent a name loss, on the grounds that one upload's Tracks section was byte-identical to Songsterr's own export and the names still went. The four measurements in section 2 refute it. A Tracks section compared after parsing is byte-identical whether or not the source carried CDATA, because the parser has already discarded the wrapper. The comparison could not see the variable it needed to see.

Zomby Woof preflight also refuses on GRID for a build that is provably timing-neutral, so two of the gate's refusals now have documented false-positive conditions.

7. The work that is not about ghosts

The ghost sweep is one thread. These ran alongside it and are recorded here so the board is complete.

Watermelon In Easter Hay, two repairs waiting on the queue

The foot-pedal hi-hat is missing. The isolated hat stem gives a periodic subsequence at 1.0800 s, equal to the GP quarter at the snare-anchored tempo of 1.0842 s, with 178 of 505 grid slots carrying a hat onset at 35% against a 26% chance baseline. Brandon's representative chicks at 181.93, 183.03, 184.10 and 185.19 appear in the detection at 181.92, 183.00, 184.09 and 185.17. The tab notates six Pedal Hi-Hat events, so roughly 172 are missing.

The snare rebound ghosts are missing. The tab carries 193 snare events with zero ghost flags. Brandon measured 202 principal snare attacks, of which 196 carry a substantially quieter follow-up within 250 ms and 178 carry two, the first at a median near 92 ms and the second near 156 ms. An independent run agreed on the lags at 70 ms and 134 ms and over-detected principals at 411 against 202, so his counts govern. These belong on the snare lane, not on Ride (middle), and whether the rebounds are stick bounce or room is not settled by a separated stem.

Bar 54 is the only tuplet bar in the tab, a 1/6 quarter-note duration inside 5/4, and its quintuplet fill needs checking against the Drumnet chart before anything is written.

The ride question, where it actually stands

Settled: the 193 post-snare Ride (middle) ghosts are not misplaced snare rebounds. The tab places them at a rigid 271 ms after the snare, one GP sixteenth, with the 10th and 90th percentiles both at 271 ms, while the measured rebounds land at 70 to 92 ms and 134 to 156 ms.

Unsettled: the other 1,123 ride ghosts. The cymbals stem gives a 41% chance baseline that swallows the signal, ghosts sit at 47% and plain at 55%, and the neither-detected rate is 50.3% for ghosts against 46.8% for plain. Those are close enough that the detector is missing real strokes. A second machine-readable transcription cannot arbitrate either, because it runs 230 bars of 4/4 plus 48 of 2/4 against the tab's 53 of 4/4 and 52 of 5/4, which makes a bar-keyed diff meaningless. No deletion revision may ship against these.

Carolina Hard-Core Ecstasy, two findings separate from the ghosts

About 16.5 bars are unnotated. The tab totals 313.45s across 102 bars, with tempo automations of 79 BPM from bar 0, 71 at bar 98, 66 at bar 100 and 45 at bar 101, against a 360.0s record. The shortfall of 46.55s is 12.93%. Filling 360s at the tab's own 79 BPM needs 118.5 bars against the 102 present. Raising the tempo to Sloots's 86.4 widens the gap by 30.12s, so this is missing material and not a tempo error.

The opening is written straight where a professional transcription writes triplets. Sloots marks triplet brackets in both staves of the opening. The tab carries 23 tuplet events in the whole 102-bar file, 6 in Drums at bar 42 and 17 in Trombone at bar 68, and bars 1 to 8 hold 505 events across all nine tracks with zero tuplets. Aligning his boxed rehearsal numbers to tab bars assumes a shared downbeat origin, so that has to be landmark-anchored against the recording before any rhythm changes.

Keep It Greasey, the tempo and anchor work

Printed sources give quarter = 134. A local sixteenth-grid sweep over the 3:18 to 3:58 solo window peaks at 143.5 with 71.1% against a 45.4% chance floor, with 135.0 at 69.7%. The grid lock is established and the exact tempo is not, so 134 must not be quoted as confirmed by audio.

Tempo-map drift is small and measured at 11 ms per bar, 0.55%. Where measure 102 begins in the audio is the binding limit, with anchor uncertainty at 1,928 ms, which is 0.91 of a 19/16 bar, across three independent families: drums at +1.176s, bass at +0.410s and keys at -0.752s.

Two corrections came from a second reader. The denominator was wrong: four sources were evaluated and three sparse parts were skipped, and counting the skipped three as failures turned missing estimates into negative observations. The second-separator result also needs a masking control, because the comparison stem is whole-kit drums, so a genuine quiet stroke masked by a louder kit event in the same 30 ms window scores as uncorroborated and is indistinguishable from an artifact.

Nine printed charts are still unapplied

Across seven songs, ranked and archived: Dong Work for Yuda s768189 (Sweeney), Apostrophe' s35873 (Bishop on Jim Gordon), Watermelon s35881 (Drumnet), Central Scrutinizer s3554079 (Bishop), Don't Eat the Yellow Snow s35866 (Bishop), Andy s21495 (Ryan Brown), and Keep It Greasey s604777 with three sources for verification only. Apostrophe' holds four tracks all on instrumentId 29, so a drum staff has to be added before a drum chart can be applied at all.

8. The seventh instance, committed while writing this update

This page documents six occasions where a check enumerated what it expected and read absence as evidence. Building this update produced a seventh.

To measure the name loss I read api/meta/<songId> and counted tracks whose title was empty. Every one of twenty tabs returned 100% empty, including Apostrophe', which this work has never touched, and Keep It Greasey, whose strip was blocked before it could publish.

A result that says every tab on the platform lost every name is a result about the instrument. The track object has no title key at all. The field is name. The real schema is difficulty, hash, instrument, instrumentId, isEmpty, isVocalTrack, name, tuning, views.

The tell was the control. Apostrophe' and Keep It Greasey were in the sweep as untouched comparisons, and both scored identical to the damaged tabs. Had the query covered only the tabs under suspicion, the number would have looked plausible and would have been published.

This is the same failure as grepping <Ghost> in GPIF, as reading a missing fret as a rest, as reading a missing dynamic key as an absent marking, and as the hardcoded allowlist in noteloss_audit.py. Assume a field name, find nothing, and report nothing as a finding.

9. What happened

Songsterr draws a ghost note as a notehead inside round brackets. That is the platform's own ghost notation and it carries real musical information, the quiet stroke a drummer plays between the loud ones. A sweep run on 2026-08-29 and 2026-08-30 read those brackets as a defect and removed them across twenty Zappa tabs.

10. The original ledger: 16 tabs, 3,842 ghost notes

Each revision below was confirmed as the current public revision by reading api/meta/<songId> on 2026-09-05.

Rejected by moderators: 2 tabs, 682 ghost notes

Superseded: 2 tabs, 41 ghost notes

Reviewed and not blocked, and another author's revision became public first. The stripping never reached a reader.

Flagged, never submitted: 4 tabs, 19 ghost notes

Root cause, three layers

1. A house rule applied outside its scope. The staccato-dot-at-velocity-31 rule comes from stems-to-guitar-pro-drums, which builds a tab from an isolated drum stem. The sweep carried it onto community tabs written by other people.
2. The wrong element name hid the evidence. GPIF spells a ghost <AntiAccent>Normal</AntiAccent>. A search for <Ghost> returns zero in every file, and that zero was read as proof the ghosts were absent.
3. The audit measured note counts and missed the loss. The preservation check compared pitch, rhythm and position across 120,579 beat events and found zero differences. Articulation sat outside what it measured.

Which lane each flag actually sat on

The count above says 3,842 ghost notes. Counting them by MIDI lane in each tab's pre-sweep revision says something narrower. A ghost note is a quiet snare stroke between the loud ones. A ghost flag on a ride cymbal is a different thing and needs a different decision.

Checked against the printed charts

The Zomby Woof repair was re-checked against the four transcriptions held on disk. Every figure below was measured off the scan rather than read by eye.

The marks below Ryan Brown's staff are pedal hi-hat

Brown's DRUM Magazine page carries small crosses one full space below the bottom staff line. Staff lines were detected at a 12px spacing, then every mark sitting under the bottom line was measured.

Eight marks, pixel-identical, across three of the four systems. A five-line drum staff puts the pedal hi-hat in exactly that position. Drumnet's bar 1 carries the same three marks at the same height, drawn in its own dialect with o and + above the staff for the hands. Two professional charts agree.

Brown's bar 1, glyph by glyph

Connected components with the staff lines erased, measured across x = 105 to 436.

The three accents land within 2px of the three open hi-hats. The three pedal hi-hats land within 2px of the three snares. The rest is the alignment anchor, since the tab also rests at slot 4 and then runs snare, hat, snare, hat, snare across slots 5 to 9.

Bar 5 accents sit on notes one and three, and Songsterr cannot draw them

Brown beams bar 5 as five quintuplet groups of five. The measured accent glyphs, width 18 and ink 65, sit above notes one and three of every group. The tab marks note two and nothing else.

What the shipped build changed, and what it left alone

Evidence, and how to reproduce each claim

Two statements on an earlier version of this page ran ahead of what had actually been measured. Both were re-run properly, and both held.

Source scans are in ~/Projects/_outputs/zappa-drum-sources/, one folder per publisher, each with a SOURCE.md giving the URL, pixel size, byte count and decode check. The measurement crops are in s412162/audit/. Revision hashes come from the image field of each revision record, which is what addresses that revision's part JSON on CloudFront.

Status, in three separate kinds

An earlier version of this page said Zomby Woof "is repaired". That collapsed three states into one. They are kept apart here.

What is left

• Sixteen live tabs need a restoring revision, one tab per pass, each verified against the pre-sweep revision rather than against the stripped one.
• Restore every attribute the sweep touched, not the ghost flag alone. The sweep wrote a staccato dot and set velocity to 31 as it cleared each ghost, so a restoring pass has to put back the ghost flag, remove the substituted dot, and return the velocity, all measured against the pre-sweep baseline.
• Keep other people's later edits. Several of these tabs have had legitimate revisions since. A restoring pass rebases onto the current public revision and changes only the articulation fields.
• Watermelon In Easter Hay is the largest single loss at 1,317 flags on the Vinnie Colaiuta staff, and its revision claimed they were ride and crash rather than ghosts, so that reading needs checking against the audio before anything is restored. Done 2026-09-06, and it came back inconclusive. 1,123 ride ghosts are unexplained.
• Verify the revision that actually becomes public, per tab, after moderation. A submission is not a repair.
• Keep It Greasey and Zoot Allures need nothing. Both stripping revisions were rejected and are blocked.

Update, 2026-09-05 18:20 UTC

Re-queried every tab after the first pass. Two findings change the work list, and the biggest restore now has a settled answer.

Need no action

Still damaged and live

Watermelon In Easter Hay, verdict

This was the one tab held back by a caveat. Its stripping revision claimed the 1,317 marks were ride and crash at forte, so not ghost notes. Both halves of that reasoning fail against the data.

The Watermelon accent grid, every slot

Every ride note on the Vinnie Colaiuta staff, counted by its sixteenth-note position within the bar. The quarter pulse is plain, everything between it is ghosted, and the split is close to absolute.

Slots 2 and 3 carry more notes than their neighbours because some bars subdivide further. The pattern holds regardless: a ride note on the pulse is written plain 407 times against 2 exceptions, and a ride note off the pulse is written ghosted 1,314 times against 1.

All sixteen restores are submitted

A parallel session uploaded every damaged tab. Watermelon went up as r8908034 and reads back from CloudFront with 1,317 ghosts, 1,316 on ride and 1 on crash, matching the pre-sweep revision exactly.

A challenge to these numbers, resolved

Another session queued a finding that this page's per-song counts disagree with the local ORIGINAL files, reporting a total of 5,322 rendered ghosts across 22 files against the 3,832 counted here. Measured properly, every number on this page holds. The gap is scope.

The drum-staff column matches this page on every row. The larger file totals count ghost noteheads on vocal, guitar, keyboard and mallet staves, which the sweep never touched and correctly left alone. Montana's 257 non-drum ghosts break down as Vocal 1/2/3 102, Frank Zappa Lead 112, George Duke 23, Ruth Underwood Bridge 14 and Tom Fowler 6, matching the 2026-08-30 audit record exactly.

Zoot Allures looked like the one row that agreed for a reason worth stating. It carries zero non-drum ghosts, so its two totals coincide. That coincidence is what made the other four look wrong.

Why the sweep passed its own audit

The 2026-08-31 handoff records that no notes were lost, measured across 20 folders and 120,579 beat events with zero pitch, rhythm or position differences. That statement is true. It was also produced by an instrument that could not see the thing being destroyed.

The sixth instance, caught by Brandon

Two sessions agreed that Songsterr's drum note schema has no accent field, so a drummer's accents could not be notated. Brandon opened the editor and said he could see the accent option. He was right, and both of us had made the same error this page spends its length documenting.