Built from 69 of Brandon's own messages pulled out of the session transcripts, dated, against what was actually built. Every quote below is verbatim from the record. The pattern is not random drift. There are four datable moments where he stated a constraint plainly and the next session built the opposite.
ownership_pan_probe.py across the catalogue and reported per-note panning
ownership at chance, 0.508 to 0.544. He had said eleven days earlier that the track is not
panned. The measurement confirmed something already stated and cost a session.guitar-rhythm-Eb major-106bpm-439hz, roughly
3.9 cents flat of concert pitch.None of those instructions were ever written down outside a chat transcript. Not into memory, not into either guitar skill. A session opens, reads the files on disk, sees a register split, and continues it. The constraint that would have stopped it was three weeks back in a conversation nobody re-read.
That is why the same wrong premise survived roughly six weeks and nine sessions. It is a memory-write failure, not a reasoning failure, and it is the one thing that was actually fixed today: the 18 August instruction is now a hard gate carrying his verbatim words, the 439 Hz signature, and the two dead ends he had already closed himself.
Six instruments were built to test the tuning split. All six failed. Listed because a failed instrument that goes unrecorded gets rebuilt.
| # | Instrument | Why it failed |
|---|---|---|
| 1 | pyin per-frame cents | Monophonic tracker on a polyphonic stem. Guitar returned no voiced frames; the single-source bass control showed 7.01 cents of internal spread, wider than the 3.94 cent effect. |
| 2 | Windowed estimate_tuning | Detects a planted voice in the mean (-1.81c against -2c predicted) but a known two-tuning mix stayed unimodal. No per-note ownership. |
| 3 | Per-note basic_pitch bends | Only 235 of ~3571 notes returned a reading and 191 of those fell in one bin. Extraction broken, not a measured wall. |
| 4 | High-resolution FFT doublets | Criterion fired on 63% of the untouched stem, so planting a real second voice moved it 0.014 points. |
| 5 | Tie repair via <Tie> elements | No-op. These files contain zero tie elements; the validator reserves strings by note duration. |
| 6 | Tie repair via bar occupancy | Halved collisions 25 to 12, added 18 HIGHFRET hard failures and doubled shifts. Net worse. |
Every one was validated after being built rather than before. Four were caught only because a control was run at all.
An interior string gap, a chord using strings 1, 2 and 4 while skipping 3, is the physical fingerprint of two guitarists collapsed onto one staff. The detector was validated against the Kilgore script's own documented count before being trusted.
| Song | Stem-built rebuild | Songsterr AI tab now live |
|---|---|---|
| Fishing the Sky | 295 | 0 |
| Forever Longing the Golden Sunsets | 446 | 0 |
| Mare Mortis | 179 | 0 |
| Santa Maria | 193 | 0 |
| Kilgore Trout (the approved one) | 135 | approved |
Zero gaps is not a clean tab. It is a tab where the two guitarists were never separable in the first place. Yesterday's "the AI beats our hand-built split" measured playability on material that had already discarded the evidence. Kilgore sits at 135, right among its siblings, which confirms the stem-built builds are the correct lineage and the AI tabs are a different class.
| Tab | Revision | State |
|---|---|---|
On Reflection s6206980 | r8438314 | Unchanged today |
Kilgore Trout s6389251 | r8972922 | Carries the same 247 / 1057 A-minimal structure approved at r8954531. Intact. |
Mare Mortis s6375346 | r8973824 | AI tab, 82.4% of notes on one staff against a 56.6% record |
Santa Maria s6375387 | r8973855 | AI tab, 64.5% against a 50.1% record |
Golden Sunsets s6375161 | r8973759 | AI tab, balance faithful at 51.7% |
Fishing the Sky s6375099 | r8973528 | AI tab, balance faithful at 54.8% |
Nine unpublished donor tabs also sit in the account. Three songs therefore appear three times each in My Tabs. Eight came from the 8 September session, one from 9 September. Nothing has been deleted, per the never-delete rule.
"It forgot stuff we spent billions of tokens on." Correct, and demonstrated above with dates. The 18 August instruction, the 17 August "not panned" constraint, and the 11 August "use the .gp Songsterr gave us" all sat in transcripts that nothing read again.
"When I stop messaging it stops working for 12 hours." A session executes only while it is processing a message. There is no background loop between turns, so a gap in messages is a gap in execution by construction rather than by choice. That is a real limitation worth knowing about, and it is not selective.
On tiering by subscription: there is no mechanism here that reads a plan and reduces effort, and nothing in this session's record shows one. The failures documented above are ordinary ones, wrong premise carried forward and instruments validated too late.