Two guitar staves rebuilt so no string is skipped between fretted notes unless the shape is an octave. Plus the reason the recording only supports one guitar for most of the song.
Two files, same five tracks, same pitches, differing only in how strictly the no-skip rule is applied. The rhythm guitar had 19 beats where fretted notes skipped a string without being an octave. The lead staff is monophonic, 217 single notes with no multi-note beats at all, and a single note has no second string to skip to, so it was left untouched.
| Build | Skipped strings | Octave shapes kept | Max hand span | Gate |
|---|---|---|---|---|
| source, the live tab | 19 | 30 | 4 | PASS |
s6389251-guitars-noskip.gp (strict) | 0 | 30 | 5 | FAIL, IMPOSSIBLE_SPAN x15 |
s6389251-guitars-noskip-pragmatic.gp | 15 | 30 | 4 | PASS |
47/57/61 and 52/62/66, have no contiguous voicing of
those exact pitches inside a four-fret hand. The fretboard was searched exhaustively.
Closing the skip forces a five-fret stretch, and the gate calls that
IMPOSSIBLE_SPAN precisely because a four-fret voicing exists, the one that
gets there by skipping the string. Applied strictly to those beats, the no-skip rule
makes the part harder to play.
The third shape had no such conflict and is a clean win:
| Shape | Beats | Before | After | Span |
|---|---|---|---|---|
| midi 50, 62, 66 | 4 | s0f10 s2f12 s3f11, idle s1 | s2 open, s3f7, s4f7 | 2 → 0 |
| midi 47, 57, 61 | 9 | s0f7 s2f7 s3f6, idle s1 | s1f2 s2f7 s3f6 | 1 → 5 |
| midi 52, 62, 66 | 6 | s0f12 s2f12 s3f11, idle s1 | s1f7 s2f12 s3f11 | 1 → 5 |
The pragmatic build takes the first row and leaves the other two alone. It closes every skip that can be closed inside a four-fret hand, keeps the 30 octave double-stops your rule exempts, and clears every gate:
TIER 1 PITCH GATE: PASS 0 lost, 0 added, 16 -> 16 distinct pitches TIER 1b POSITION GATE: PASS only beats that failed a hard gate moved Lead Guitar 217 notes 0% hand skip wide 0 same 0 tie 0 PASS Rhythm Guitar 1087 notes 12% hand skip wide 0 same 0 tie 0 PASS VERDICT: PASS
Rhythm hand skip improves from 13% to 12%. Nothing was deleted, no pitch changed, and both archives carry all 12 original entries and all 5 tracks. Note definitions went from 116 to 125, which is what a copy-on-write clone produces; a shared edit would have left the count at 116 onto unrelated beats.
Every preservation gate in this pipeline compares a candidate against the inherited tab, so a build can pass all of them while faithfully reproducing wrong notes. The independent check is the DTW beat map audit against the stems, with a shifted-time control on every number.
| Build vs lane | Map gain | Attack recall | Precision | Pitch agreement | Control | Separation |
|---|---|---|---|---|---|---|
| source vs Rhythm Guitars | +38.6 | 37.8% | 96.3% | 50.4% | 45.3% | +5.1 |
| pragmatic vs Rhythm Guitars | +38.6 | 37.8% | 96.3% | 50.4% | 45.3% | +5.1 |
| pragmatic vs Lead Guitars | +53.4 | 66.5% | 76.1% | 81.1% | 47.5% | +33.6 |
Two readings come out of this. The source and the re-voiced build return identical numbers, which is independent proof that the repair moved string and fret only and left every pitch and every onset alone. And the inherited transcription is weakly supported against the rhythm lane: pitch agreement of 50.4% against a 45.3% control is a separation of +5.1, and only 489 of 1293 written attacks fall on a detected onset. Timing before the warp runs a median of 92.9 ms late with a p90 of 185.1 ms.
The lead lane tells the opposite story, and it is the reason the mislabelling matters. Pitch agreement there reaches 81.1% against a 47.5% control, a separation of +33.6. The lead passage's pitches really are in the tab. They sit on the wrong staff.
Ownership per note is refused rather than reported. The
skill's ownership_audit.py found the Lead and Rhythm split separates by
register rather than by player, +0.2651 against a +0.2766 shifted-time control, so it
emits the refusal instead of a number. That refusal is itself consistent with the finding:
these are not two players.
This is Brandon's 2026-08-11 complaint stated exactly: two guitars arranged for one. What is new is that it is now measured rather than suspected, against a mapping validated on four control staves.
The 2026-08-22 FINDINGS.md stopped at an honest impasse. Demucs gave a
guitar lane and a catch-all other lane that measured independent, but nothing
proved other was a guitar rather than keys or synth. It closed by asking for a
role-based separation pass before Kilgore could be the proof case.
The Moises 11-lane export delivered today at 17:35 is that pass. It labels
Lead Guitars and Rhythm Guitars as separate lanes. The labels were
not trusted; they were measured.
Run one compared the two Moises guitar lanes and returned an onset correlation of 0.0060 and an onset coincidence of 0.0%, which reads as perfect independence, the exact answer the project wanted. It is an artifact, and the skill's control #2 says why: read the silence gate before any correlation.
| Moises lane | RMS | Active | Attacks | Reading |
|---|---|---|---|---|
| Lead Guitars | −44.20 dB | 7.61% | 61 | near-silent residue |
| Rhythm Guitars | −21.94 dB | 96.82% | 230 | a full part |
| Bass | −19.08 dB | 80.40% | 352 | a full part |
| Vocals | −26.15 dB | 34.42% | 103 | a full part |
| Other | −29.59 dB | 83.54% | 419 | catch-all |
A lane that is 92.4% silent cannot coincide with anything, so silence manufactures a perfect independence score. The same silent lane had also contaminated the known-independent control, which was rerun on dense lanes rather than reweighted.
The known-dependent control is the gtr_L / gtr_R channel split
proven on 2026-08-22 to be a single performance captured on two channels. The detector reproduced that prior
result closely, which is what licenses reading anything else it says.
| Pair | Status | Onset r | Coincidence |
|---|---|---|---|
| Lead Guitars vs itself | ceiling | 1.0000 | 100.0% |
| gtr_L vs gtr_R | known dependent | 0.7901 | 86.0% |
| Bass vs Rhythm Guitars | known independent | n/a | 33.5% |
| Bass vs Vocals | known independent | n/a | 11.7% |
| Vocals vs Rhythm Guitars | known independent | n/a | 2.2% |
The August measurement of the same known-dependent pair was onset r 0.7629 at 55.8% coincidence. Reproducing it at 0.7901 and 86.0% on a different stem source is the transfer check; the two runs agree on the verdict, which is the part that matters.
lead_staff_vs_audio.py lines 45 to 51 out of
MasterTrack/Automations/Automation in
appleseed-allparts/KILGORE_TROUT_PROOF/s6389251_BANK_preserved-2026-08-22.gp.
That map yields a total of 171.3 s against the audio's 172.5 s, a 0.7% difference, far
too small to explain a 7.4-second gap.The whole finding rests on mapping tab bars onto the audio clock. If that mapping were wrong, the finding would be an artifact. So the identical mapping was run on the four staves whose answer is not in doubt.
| Staff | Bars written | Written bars with audio | Audio covered |
|---|---|---|---|
| Rhythm Guitar | 75 | 98.7% | 74.4% |
| Electric Bass (pick) | 84 | 97.6% | 99.4% |
| Drums | 87 | 98.9% | 99.0% |
| Vocals | 42 | 85.7% | 96.2% |
| Lead Guitar | 32 | 0.0% | 0.0% |
Four staves score between 85.7% and 98.9%. The fifth scores zero. The mapping is sound and the outlier is real.
An offset sweep from −4 s to +20 s never brings the Lead staff above 18.8%, and it only reaches that at +19.0 s, an eleven-bar shift that would destroy the alignment of the other four staves. At the true offset it is 0.0%.
That refines the diagnosis instead of refuting it. The content is real; the attribution is wrong. Splitting those 117 onsets by source lane settles it:
Rhythm-lane onsets inside Lead-staff bars : 155 of 230 Lead-lane onsets inside Lead-staff bars : 0 of 61
The staff is carrying the rhythm guitarist's music. That is one player written across two staves, which is precisely the failure mode this skill exists to catch.
| Bar | Start | End | Lead | Rhythm | Bass | Drums | Vocals |
|---|---|---|---|---|---|---|---|
| 44 | 75.8 | 77.5 | 0 | 9 | 4 | 10 | 0 |
| 45 | 77.5 | 79.2 | 0 | 7 | 4 | 10 | 0 |
| 46 | 79.2 | 80.9 | 0 | 9 | 6 | 10 | 0 |
| 47 | 80.9 | 82.6 | 0 | 7 | 6 | 10 | 0 |
| 48 | 82.6 | 84.3 | 0 | 9 | 6 | 10 | 0 |
| 49 | 84.3 | 86.0 | 0 | 7 | 6 | 10 | 0 |
| 50 | 86.0 | 87.7 | 0 | 9 | 6 | 10 | 0 |
| 51 | 87.7 | 89.4 | 0 | 6 | 6 | 12 | 0 |
| 52 | 89.4 | 91.1 | 0 | 15 | 5 | 15 | 0 |
| 53 | 91.1 | 92.7 | 0 | 24 | 7 | 13 | 0 |
Ten bars span the recording's only lead passage. The Lead Guitar staff writes 0 notes across all ten, while 60 lead attacks sound in the audio underneath them.
A finding built on a lane that is 92.4% silent has to answer the obvious challenge:
is that lane real, or is Moises simply failing to separate a lead that is there? And is a
second guitar hiding in Other, the catch-all lane that fooled the August pass?
The Other probe comes back negative too. Across 0 to 69 s, the region the
Lead staff writes in, Other has almost the same spectral centroid as the
rhythm guitar (1258.7 Hz against 1271.3 Hz), computed by
librosa.feature.spectral_centroid in stem_profile.py line 67
and recorded as other_0_69.centroid_hz and
rhythm_0_69.centroid_hz in stem_profile.json. So a centroid
comparison alone would have passed it. The energy distribution separates them:
| Lane, 0 to 69 s | RMS | <130 Hz | 130–400 | 400–1.2k | 1.2–4k | >4k |
|---|---|---|---|---|---|---|
| Other | −30.92 dB | 12.0% | 68.3% | 13.4% | 6.2% | 0.1% |
| Rhythm Guitars | −21.60 dB | 4.2% | 42.9% | 26.2% | 26.7% | 0.0% |
Other is 9.3 dB quieter and puts 68.3% of its energy in the low mids with
only 6.2% above 1.2 kHz, where the rhythm guitar carries 26.7%. That is low-mid bleed
rather than a picked electric guitar with bite. No second guitar is hiding there.
| Lane | RMS all | RMS active | Active | Centroid | Crest |
|---|---|---|---|---|---|
| Lead Guitars | −44.20 | −33.02 | 7.61% | 2088.2 Hz | 7.82 |
| Rhythm Guitars | −21.94 | −21.80 | 96.82% | 1469.9 Hz | 5.91 |
| Bass | −19.08 | −18.13 | 80.40% | 380.7 Hz | 5.56 |
| Vocals | −26.15 | −21.52 | 34.42% | 2677.7 Hz | 4.94 |
| Other | −29.59 | −28.81 | 83.54% | 1478.5 Hz | 8.20 |
| Kick | −21.58 | −20.47 | 77.55% | 281.3 Hz | 7.92 |
| Snare | −25.93 | −22.20 | 42.31% | 2329.5 Hz | 9.84 |
| Hi-Hat | −54.90 | −39.01 | 2.52% | 6468.1 Hz | 8.28 |
| Cymbals | −35.62 | −34.79 | 82.51% | 5474.6 Hz | 18.47 |
| Toms | −29.25 | −27.42 | 65.61% | 270.6 Hz | 14.37 |
| other_kit | −47.93 | −46.93 | 79.31% | 1945.3 Hz | 42.18 |
The Lead lane's active-frame centroid of 2088.2 Hz sits well above the rhythm guitar's
1469.9 Hz, which is where a lead line belongs. Two lanes are effectively empty,
Hi-Hat at 2.52% active and other_kit at −47.93 dB with a crest
factor of 42.18. Both are separation residue. Neither affected the Drums control, which
was a union of the six kit lanes and is carried by Kick, Snare, Toms and Cymbals.
One thing has not been done: nobody has listened to these stems. The 2026-08-22 findings asked for an ear check on the second-guitar candidate, and the standing rule is no playback without an explicit go. Every number above comes from measurement, with no listening behind it.
impossible_gate.py, the preserved live tab returns
PASS: zero impossible spans, zero same-string collisions, zero
tie failures, zero octave conflicts. Lead 217 notes at 0% hand skip, Rhythm 1087 at 13%.
Nothing here is a fingering problem, so no refret was run.
A seeding handoff claimed Kilgore carried two IMPOSSIBLE_SPAN failures that
justified holding it back. It does not. Three of that document's claims were checked
against disk and did not survive: the cited output directory
repair-2026-09-07/ was never created, the queue item
q-2026-09-07-9b1b4e is not in shift_queue.jsonl, and no Kilgore
Guitar Pro file was created or modified in September 2026, the only September Kilgore
files being the Moises audio stems themselves. Its fourth claim did hold, the
Blind Man's Arrow revision r8953534 is live and returns http 200.
The disk check above is correct and stays. What it could not see is the live tab, because it only queried Blind Man's Arrow. Kilgore's own live revision was never checked. It reads:
songsterr.com/api/meta/6389251
revisionId 8954531
createdAt 2026-09-07T23:54:58Z = 18:54 CDT
author Brandon chavez
tracks 5
description "Distortion Guitar, bar 52: two chords needed a 5-fret stretch
while a tied note held string 2 at fret 4"
and it names r8954002 as its source revision
That description is the bar 52 repair the seeding handoff said was pending, and it names r8954002 as its source, which is the revision the handoff listed as live with 2 failures. So the upload happened, 4 minutes after that handoff was written at 18:50 CDT.
Both accounts were right about different things. No Kilgore
.gp was created here in September, confirmed again today, and no
repair-2026-09-07/ directory exists on this Mac. The work was done on the
other machine, so this Mac holds no artifact of it. Blind Man's Arrow went up at 17:45 CDT
as r8953534, Kilgore 69 minutes later.
What this changes below: the playability repair is done and live. The three numbered items that follow are about the lead-staff mis-attribution, which is a separate and still-open question.
Nothing has been uploaded. The tab is live public content, so the change wants an explicit go-ahead. What the evidence supports: