The Gate That Taught Itself To Quit

What Brandon said

The gate was written on 1 September to stop a chat from running two Zappa songs thin. Three amendments followed inside five days, and every one of them repaired how the lock identifies a song. Amendment one keyed it to a song id after two tabs titled Zomby Woof got edited in one session. Amendment two added a third branch for a song held by a moderator. Amendment three added a 24-hour ceiling after a 113-hour chat.

None of the three touched the clause Brandon is naming. The original rule ends with “Any request that touches a second Zappa song opens a new chat”, and it supplies a chat with two respectable-sounding exits: declare the song finished, or declare the remaining songs somebody else's problem. Both got used, in writing, more than once.

What the queue shows

Source: ~/.claude/shift_queue.jsonl, 298 records, latest-by-id wins. Session counts from 85 transcripts under ~/.claude/projects/-Users-brandonchavez/.

When the Zappa backlog was created

The item at the front of the queue

queue.py next has returned the same record since 31 August:

The files exist and were verified lane for lane against the iMac census:

~/Projects/_outputs/zappa-handoff-lie-audit/

The stated blockers were both wrong, measured 6 September

The queue item claimed two blockers. Under the new rule the first act is to work the item rather than repeat its label, so both were tested.

What preflight actually reports is a different pair of faults:

The four repairs

Repair 1 · the lock governs focus and carries zero authority to stop

The single-song lock decides which song a turn works on. It decides nothing about whether the chat keeps working. Every stop in this lane now needs a reason that stands on its own with the gate unmentioned. The two sentence shapes quoted above are gate violations from 6 September forward.

Repair 2 · a finished song RE-LOCKS inside the same chat

One song at a time survives intact. One song per CHAT is retired. The focus discipline was always about where a turn puts its attention. Splitting it across chats bought nothing and cost 77 queued items.

Repair 3 · queueing a Zappa song defers it by hours

Queueing keeps one legal purpose, ordering the work inside a chat. A queued Zappa item gets worked in the same chat as soon as the current id closes. It stops being a way to move work into a chat that nobody opens.

Repair 4 · the 24-hour ceiling stays, and it hands over

Amendment three's ceiling of 24 hours or 400 assistant turns answered a 113-hour chat and stays in force. It fires as a handover. The handoff names the LOCKED id, the finish state, and the exact next queue id, so the successor chat re-locks and resumes inside its first response. A ceiling breach never empties the queue.

Precedence, settled

Branch three, worked rather than described

The new rule says a chat held by a moderator works evidence on the same song. That ran here, and it turned up two more errors in the queue item plus a measurable defect in the tab.

A new independent source, fetched today

Sloots's 1975 transcription was missing from the local mirror and came back http 200 from zappa-analysis.com: the page, two notation scans at 875×362 and 828×449, and a MIDI. Parsed from that MIDI: 4/4 throughout, 86.4 BPM, 8 bars, 399 note-ons. It carries keyboard, bass and melody, so it votes on the opening 8 bars alone and says nothing about the kick lane.

The two agree on meter. The tab runs 8.56% slow against Sloots's MIDI on the opening, which is 7.4 BPM out of 86.4.

Glyph provenance, read at native resolution

The 86.4 BPM is the MIDI's tempo meta event and nothing else. The score itself was checked with a 1:1 crop of carolina-hard-core-ecstasy-1975a.jpg (875×362 native), cut at full pixel resolution and upscaled only for viewing.

The top system carries a boxed rehearsal number 1, an 8va, treble clef, a key signature of one sharp, and 4/4. The bass staff repeats one sharp and 4/4.

Restore verified by set difference, not by count

Both revision descriptions say “144”. A count is not proof, and element counts mislead here, so the check walks played events. The flag is <AntiAccent>, a child of <Note>, never a Property, which is why a Properties-only audit reads zero ghosts on a tab that renders them.

GPIF Notes are shared definitions, so one element carrying AntiAccent can serve 144 beats. Reading the element column literally would say the restore invented 143 ghosts.

Triplet disagreement, read at native resolution from scan b

carolina-hard-core-ecstasy-1975b.jpg is 828×449 native. Two 1:1 bands were cut at full pixel resolution and upscaled only for viewing. Band 1 carries a boxed rehearsal 5, an 8va, treble clef, one sharp, no restated time signature, and triplet brackets marked 3. Band 2 carries triplet brackets marked 3 in both clefs across three of four visible bars. Neither band carries a metronome mark.

The whole tab carries 23 triplet events and none of them are in the opening. Bars 1-8 hold 505 events across all 9 tracks, and 0 of them are tuplets. The tab writes the opening straight while Sloots writes it with triplets.

The 13.92% grid gap is missing material

Against a 360.0 s record that is a 46.55 s shortfall, 12.93% of 360. Filling 360 s at the tab's own 79 BPM needs 118.5 bars against the 102 it carries, so about 16.5 bars of content are unnotated. Raising the tempo toward 86.4 makes it worse, since 102 bars at 86.4 runs 283.33 s and widens the gap by 30.12 s.

Verify it yourself