Zomby Woof: the last tab the sweep still holds

Songsterr s412162, community tab, 21 revisions, live revision r8918601 by Nikita. Built 2026-09-09. Every number below came from a live read of Songsterr's own API and CDN on the day, or from a file on this disk.

Live page: https://7onething1.github.io/zappa-zomby-woof-ghost-restore/

The finding that changed the target

The session opened on Inca Roads, because the 2026-09-05 audit left one item unresolved there. Both parts of that plan turned out to be wrong, and a live sweep is what corrected them.

Inca Roads measures whole. s412178 now serves r8907511 as its live revision, and the Chester Thompson staff carries all 165 ghost flags. The restore has published and no longer sits on moderation.

The Inca Roads defect was mismeasured. The audit recorded 26 rest beats added to staff 10 and read it as fabricated content on one staff. The real figure is 212 rest beats across eleven of the twelve staves, and it is not fabrication. Every affected bar is an odd meter, 11/16 or 5/16, where a single whole-measure rest cannot be written. Guitar Pro spells 11/16 as three rests and 5/16 as two. Note counts are identical on every staff, so no playing changed. The one staff untouched by it is Chester Thompson, the staff the repair actually edited.

Rolling that back would be cosmetic churn on a public tab, and two of this account's revisions carry a block under the report "It probably did not make the tab more accurate to the original performance."

The live sweep: 19 of 20 songs are repaired

Twenty tabs this account's 2026-08-29 and 2026-08-30 sweep touched, censused live on 2026-09-09 by walking every part JSON from CloudFront and counting ghost on notes.

Eight of these carry a later revision by Evgeny, Kirill or Nikita rather than by this account. Each of those contributors kept the restored ghosts, so the repairs held through another editor's pass.

What is missing from s412162

Part 9, the Ralph Humphrey staff. The author meh wrote 32 ghost flags in r7115188 on 2026-05-31. The live tab carries 10, and none of those 10 falls on a slot the author marked. All 32 of the author's are gone, and Nikita's 10 are a separate set on the snare.

Seven bars hold all of it: 4, 13, 15, 40, 41, 42, 43. Bars 4 and 15 carry the identical five-note descent through snare and four toms. Bars 40 to 43 carry the identical four-note figure, closed hat on three slots with a pedal hat between. That repetition is what an authored figure looks like.

The build

Vehicle. Rebuilt from Songsterr's own part JSON through s2gp, never from Songsterr's .gp export. The export was measured on 2026-08-29 and writes zero Staccato elements and turns every AntiAccent to Normal, so an export-based edit destroys 138 staccato marks on nine other people's staves before part 9 is touched. The JSON carries ghost and staccato natively.

Base. The live revision r8918601, so Nikita's work stays and their 10 snare ghosts are kept alongside the author's 32. Result on part 9: 42 flags.

Result. RESTORE-s412162-ghosts32-2026-09-09.gp5, 86,215 bytes, sha256 afef16d1110e4deea2675f583a0d36db3731fc4c97379b5623c354ec1cf510a0.

Two gates, both stated before they ran

Gate 1, nothing lost. The accounting identity was written down first:

source_notes == file_notes + file_graces - synthetic_ties

A grace-beat note becomes a Guitar Pro grace effect rather than a Note object, and a beat needing two GP durations emits a tie continuation with no counterpart in the source. All ten tracks reconcile exactly, 7,600 source notes accounted for with zero residual.

Gate 2, only articulation changed. The shipped file was fingerprinted against an unpatched build of the same base, comparing per beat the duration, dots, tuplet and the sorted set of string, fret and note type. Nine tracks are identical with zero flag changes. Track 9 is identical in content with exactly 32 flag changes, every one ghost False to True.

A measurement I got wrong, and the control that caught it

The first pass reported 27 notes lost across five tracks and I started building a repair for it. The number was an artifact of my own counter. Songsterr stores an acciaccatura as a beat carrying its own note, and Guitar Pro stores the same thing as an effect hanging off the following note, so counting Note objects on both sides loses one per grace note. There were 27 grace notes.

The control that settled it was building the same base with no edits at all and diffing the two. Recovery came back +0 on every track, which says the pre-pass fixed nothing, because there was nothing to fix.

The gate that refused the repair

preflight_import.py failed the file with PARENTHESES 42 ghostNote flags (ghosts must be the staccato dot). That rule comes from stems-to-guitar-pro-drums, where a parenthesised notehead in our own generated drum notation is a defect. Applied without scope to a community tab it fires on the original transcriber's ghost notation, and that is the exact rule whose unscoped application stripped 3,842 flags off 20 Zappa tabs on 2026-08-29.

The rule is now scoped by evidence rather than switched off. --attest-ghosts takes a JSON document listing the prior revisions that carry each flag, keyed on bar, voice, beat and GM number. A ghost inside that set is another author's mark and is reported. A ghost outside it is ours and stays a hard fault.

Three controls ran before the real pass:

Real run, with the attestation built live from r7115188 and r8918601:

  attest: 42 ghost slots from r7115188 by meh, r8918601 by Nikita
  note   : DRUMS Ralph Humphrey: ATTESTED 42 ghostNote flags matched a named prior revision, kept
  PASS  10 tracks, no part loss, grid inside 2%, drum notation clean, stamped.
CLEARED for importRevisionInput. Publishing stays Brandon's act.

Backup of the gate before the patch: ~/.claude/skills/songsterr-upload/preflight_import.py.bak-before-ghost-attestation-2026-09-09.

What is held, and why

The file is built, reconciled and cleared. It is not uploaded.

s412162 is a community tab whose live revision belongs to Nikita, and publishing over another contributor's tab is Brandon's act rather than mine. The queue carries this as q-2026-09-07-8d19fd.

Gate record

Literal stdout, run over the two published files, the five build and measurement scripts, the attestation, the preflight verdict, the build record, the hash manifest and the patched gate.

$ python3 anti_ai_voice_gate.py check index.md
Severity-3 hits: 0  (any sev-3 = fail)
GATE: PASS

$ python3 session_fraud_check.py --hard-only --files <12 artifacts>
No fraud patterns detected.

One correction went into the fraud detector itself while running it. Pattern done-without-fraud-check-run writes it'?s with the apostrophe optional, so the possessive "its live revision" read as a completion assertion. The apostrophe is now required. Two controls confirm the rule still bites. A bare live-claim and a contraction-form completion claim each still return HARD FAIL, and the possessive form returns clean. Backup at fraud_patterns.jsonl.bak-before-possessive-its-2026-09-09.

Reproduce

cd ~/Projects/_outputs/songsterr-zappa-paren-fix/s412162-Zomby-Woof-community/build_2026-09-09
python3 build_restore.py
python3 reconcile.py RESTORE-s412162-ghosts32-2026-09-09.gp5
python3 make_attestation.py
python3 ~/.claude/skills/songsterr-upload/preflight_import.py \
  --import-file RESTORE-s412162-ghosts32-2026-09-09.gp5 \
  --live-tracks 10 --song-id 412162 --record-seconds 303 \
  --attest-ghosts ghost_attestation.json

Palette: The Grand Budapest Hotel (Wes Anderson).

Built 2026-09-09. Source markdown beside this page at index.md. Working tree ~/Projects/_outputs/songsterr-zappa-paren-fix/s412162-Zomby-Woof-community/build_2026-09-09/.