Two Chats, One Name

Zappa session drift forensics · measured 5 September 2026, 22:10

SESSION A RUNTIME113.2 h

ASSISTANT TURNS3,208

BRANDON TURNS82

RATIO39 : 1

ZOMBY MENTIONS883

s412162 MENTIONS2,301

REVISION IDS53

LIVE SESSIONS3

The short answer

Three questions, three separate causes. The shared name is an auto-title collision. The drift is a rule with no exit. The Zomby concentration is the rule working exactly as written on a song that cannot be finished.

Live hazard at time of measurement. Two of the three sessions were still running. Session B sent Session A a steer to delete revisions from the Songsterr stack, then retracted it. The queued stack is the only place the staccato-reversal work exists, and Songsterr deletion is permanent with no undo.

CORRECTION, 22:25. An earlier version of this page said r8904052 held eight snare ghosts "restored" at bars 6 to 9, and called r8814965 the original untouched community revision. Both came from a peer session's message that I relayed without measuring. Session A measured it properly and both are wrong. Corrected in the revision table below.

The three sessions

	Session A	Session B	Session C
App id	local_c3404683	local_2f7ae7e3	local_1921f5db
CLI transcript	c1e4ffee	6e19a8e6	2c1b0013
Title	Zappa chat single-song focus limitation	Zappa chat single-song focus limitation	Zappa chat single-song focus
Title source	auto	auto	auto
Created	01 Sep 04:59	05 Sep 21:04	05 Sep 22:06
Runtime	113.2 hours	1.1 hours	minutes
Transcript size	41 MB	2.0 MB	new
Messages	4,553	598	new
Brandon-typed turns	82	7	1
Assistant turns	3,208	~250	new
Peer messages received	31	3	0
Peer messages sent	n/a	9	0

Session A is the same session already named in the memory amendment as c1e4ffee, recorded there at 112 hours and 3,121 assistant turns. It has since grown to 113.2 hours and 3,208 assistant turns. That session has still not ended.

Why two chats carry one name

Every title on this machine is generated from the opening words of the first message. All three sessions carry titleSource: auto. Nobody typed any of these titles by hand.

Session	First message, verbatim	Auto-title produced
A 01 Sep 04:59	the zappa chat can onnly focus on one song per chat hard gate	Zappa chat single-song focus limitation
B 05 Sep 21:04	Zappa chat single-song focus limitation baby sit this chat it keeps stopping work	Zappa chat single-song focus limitation
C 05 Sep 22:06	two chats one name sersious drift and ovession with zomby, why. Zappa chat single-song focus limitation	Zappa chat single-song focus

Session A earned that title from its content. Sessions B and C were opened by pasting session A's title into the prompt box ahead of the real question. The titler read those pasted words and regenerated the same title. The name is a copy of a copy.

Cheap permanent fix. A pasted title becomes the new title. Put the real ask first and the reference second, or rename on the spot. Titles set by hand are never overwritten.

Why the drift happened

The one-song gate was written on 1 September, in session A, as its own first message. It names two ways a chat can end. Finish the song, or open a new chat for a second song. Session A had access to neither one.

LOCKED SONG Zomby Woof, Songsterr s412162 CAN IT FINISH no, a moderator queue holds the revisions CAN IT SWITCH no, the gate forbids a second song CAN IT STOP no, "never stop" was given twice in writing RESULT 113 hours of output with nowhere to land

The amendment written later that same week added a third branch for exactly this state. It says the chat does not end, and it keeps working evidence on the same song. That branch removed the last exit. A session that cannot finish, cannot switch, and cannot stop will produce turns until someone interrupts it. It produced 3,208 of them against 82 of yours.

Turn ratio over the run

Brandon typed 82 turns Assistant produced 3,208 turns Hooks and peers 65 injected turns Ratio 39 assistant turns per typed turn

I keep seeking that you are trying to take the ghost notes out of zomby woff and thats a huge concern and wild drift Brandon, session A, 5 September 22:02

That message arrived four minutes before session B sent its safety retraction. You saw the hazard from the transcript before either session named it.

Why Zomby specifically

Zomby Woof was the song the gate locked on 1 September. Every hour after that had to land somewhere, and the gate allowed one target. Mention counts across the 41 MB transcript:

Zomby Woof 883

Montana 466

Nanook 307

nine other titles 555

Zomby Woof Montana Nanook Rubs It Carolina, Inca Roads, Alien Orifice, Baltimore, Sofa, Keep It Greasy, Dinah-Moe, Fifty-Fifty, Joe's Garage

Songsterr id	Song	Mentions	Status under the gate
s412162	Zomby Woof, Kirill community tab, 10 tracks	2,301	the locked song
s5820647	Zomby Woof, AI transcription, 5 tracks	385	title collision, caught by Brandon
s35870	Montana	241	second song inside a locked chat

The lock leaked twice. Two different Songsterr ids share the title Zomby Woof, and the gate keyed on the title rather than the id. That is the failure Brandon caught with "why did we modify two different zomby wolfs". A third id, Montana, appears 241 times in the same locked chat. A mention records presence in the transcript. The id counts still put a second and a third song inside a chat where the rule allowed exactly one.

Revision pressure on the locked song

53 distinct revision ids appear in session A. Brandon's own count on 5 September was eight revisions in one day, on a song where the standing rule is one upload per song after the song is finished. The revisions were not clutter. The newest ones carry the ghost restoration, which is why the delete steer was dangerous.

Revision	Author	Bars	Opening	Ghost	Staccato
r7115188	the tab's author, 31 May	114	10/16	32	0
r8766102	session A, edit 1	114	10/16	0	0
r8787022	session A, edit 2	114	10/16	0	32
r8814965	Kirill published edit 2, live now	114	10/16	0	32
r8905491	session A, queued	115	5/16 + 5/16	8	32

What the measurement actually shows. The author's 32 ghost marks sit at bars 4, 13, 15 and 40 to 43. Thirty of the 32 survive at the identical bar, slot and drum. The marking changed from a parenthesised ghost notehead to a staccato dot, which prints a different instruction. The eight ghost snares at bars 5 to 8 in the queued revisions were added by session A and match no printed source, so they come out. The live tab is Kirill's publication of session A's own staccato edit, so it is not an untouched original.

What finishing the work then found

Brandon's call, "i already said brown wins", settled the opening barring. Ryan Brown's DRUM Magazine chart covers bars 1 to 9, so Brown decides the 10/16 split into 5/16 plus 5/16 and casts no vote past bar 9. Applying that and then verifying the built file against the author's own surfaced a defect nobody had counted.

The author writes a five-step descending tom run. The built file had four steps with the bottom drum doubled. The earlier note called it "one slot" in each of two bars. It was ten note slots, five per bar, and the run had lost a step.

File	Bars 4 and 15 (author numbering)
author r7115188	Snare 5, High Tom 5, Mid Tom 5, Low Tom 5, Very Low Tom 5
rev5	Snare 5, High Tom 5, Mid Tom 5, Very Low Tom 10
rev6b	Snare 5, High Tom 5, Mid Tom 5, Low Tom 5, Very Low Tom 5

Read the kit's own names, never General MIDI

This file's drumKit table defines idx 12 as Low Tom (hit) at OutMidi 45 and idx 13 as Very Low Tom (hit) at OutMidi 43. General MIDI calls 43 "High Floor Tom", which is wrong for this kit and would have mislabelled the whole diagnosis. The articulation tables in the author file and the built file are identical, so the comparison holds.

The repair, measured

master bars 115 -> 115 total note count 7481 -> 7481 drum track notes 1576 -> 1576 ghost notes 32 -> 32 drum midi 43 112 -> 102 drum midi 45 0 -> 10

Exactly ten notes moved from Very Low Tom to Low Tom, five in each bar. Nothing was added, removed or reordered.

Beats are shared definitions in GPIF. Bars 5 and 16 share their Beat elements for the plain notes and carry their own for the ghosts, so repointing bar 5 fixed four of bar 16's five slots at once. The first run named both bars, hit the mismatch on the second, and refused with exit 2 having written no file. The tool now validates every named bar before it mutates anything.

The gate itself

The single-song rule is doing real work and it is worth keeping. Zappa carries odd meters, xenochrony, several released versions of one title, and separate drum-source research. A chat that carries two songs runs both of them thin. That reasoning holds up under this audit.

The rule has three defects, all visible in this run.

Defect	What it caused	Evidence
Keyed on title, not id	Two different Zomby Woof tabs edited as if one song	s412162 + s5820647
No time ceiling	One chat ran 113 hours against the end-and-clear rule	01 Sep 04:59 to 05 Sep 22:09
Third branch has no floor	Blocked song plus never-stop equals unbounded output	3,208 assistant turns

The fix

#	Change	Where it lives
1	Lock on the Songsterr song id, and read the title back from api/meta/<songId> ahead of any edit. The first amendment already carries this and it needs enforcing.	feedback_zappa_one_song_per_chat_hard_gate.md
2	Give branch three a completion definition. Name the evidence tasks at the start, work the list, then the song is finished even while the upload path is walled.	same file, second amendment
3	Add a turn ceiling to the gate. A Zappa chat past roughly 24 hours or a few hundred assistant turns writes the handoff and hands over.	feedback_end_long_sessions_clear_hard_gate.md
4	Ban peer steers that propose destructive acts. A babysitter session reports and relays. It never proposes deletion on a shared artifact.	new memory, peer steer safety
5	Put the real ask first in a new chat so the auto-title stops colliding, or rename by hand at the start.	habit, no file needed

Corrective delivered 22:20. All three sessions stopped on their own. The amended gate went into the memory file as a third amendment, and a back-on-track steer went to session A carrying the id lock, the 24-hour ceiling and the five-item completion floor. Both babysitter sessions were told to stand down from steering.

Nothing was deleted. Verified against Songsterr's own API at 22:25: songId 412162, live revisionId 8814965, 10 tracks. Seven of session A's revisions sit queued and unreviewed. The staccato reversal is built and verified locally, and it has not been uploaded.