Provisional measurement · Phase 2 not accepted

Kilgore Trout guitar transcription

Independent verification pass run on the MacBook against the iMac's 2026-08-13 bundle. Every number below is reproduced locally from the shipped artifacts. The accuracy result is provisional because a hard gate still fails.

Machine MacBookPro · /Users/brandonchavez Bundle 27 of 27 hashes verified Date 2026-08-13

The bundle I was sent is thirteen promotions stale

The iMac's live audit had already advanced from chain49 to chain62 before this bundle arrived, and it cleared the impossible span in its Phase 3. Everything measured here reproduces its published Phase 2 numbers exactly, so the reproduction is sound and most of the conclusions are its work, not new.

One thing here is new, and it is the test the live audit says it did not run. That audit states the capo finding as a hypothesis and names the clean test. This page runs it.

NewThe clean capo test, isolated

The live audit: "This audit measured that the capo-0 Rhythm staff agrees with the audio far better. It did not isolate the capo as the cause by holding everything else fixed. The clean test is to re-render c-tied.gp's Rhythm staff at capo 0 with no other change and re-measure."

c-tied-RHYTHM-CAPO0.gp is c-tied with exactly one XML property changed, CapoFret on track index 1 from 2 to 0. Every other byte of the archive was copied through unaltered.

Build	Rhythm capo	Rhythm pitch supported	Rhythm owner agrees	Lead, the control
c-tied as shipped	2	167/677 · 24.7%	45.6%	64.2% / 54.2%
c-tied, capo 0 only	0	397/626 · 63.4%	71.9%	64.2% / 54.2%
s6389251-chain49	0	426/656 · 64.9%	71.8%	65.5% / 53.9%

The hypothesis is now a verdict. The capo is the cause.

1The swing to explain is 24.7% to 64.9%, which is 40.2 points. The capo alone delivers 38.7 of them, 96% of the total. Every other edit chain49 made is worth the remaining 1.5 points.
2Owner agreement runs 45.6% to 71.8%, which is 26.2 points. The capo alone delivers 26.3. The Rhythm staff moving to the correct stem is the same correction, not a separate one. The live audit lists those as two findings, and one cause sits under both.
3The Lead staff reads 64.2% and 54.2% in both runs. The edit touched only Rhythm and only Rhythm moved, which is the control this test needed.

Method note, stated because it matters.
librosa will not install on this Mac: llvmlite fails to build on both python 3.11 and 3.14. It is used in two places only, a 44100 to 22050 resample and a separate onset report.
The resample was substituted with scipy.signal.resample_poly at an exact 1:2 ratio, and the substitute was validated against the iMac's published chain49 numbers before any new measurement was trusted: controls 87.9% / 10.47 dB, 85.7% / 9.21 dB, 75.5% / 9.07 dB, margins 2.29 and 2.93, pitch 65.5% and 64.9%, owner 53.9% and 71.8%. Every figure reproduced exactly.
Onset detection returns an explicit not-computed marker rather than a fabricated number. It feeds none of the pitch results.

Tier 1 / 1bSource fidelity, measured against c-tied

The position gate is the one that separates a repair from a rewrite. It asks whether a note kept the string and fret the source gave it, unless that note was the thing being repaired.

Build	Hard profile	Pitch	Position	Notes moved	Gate verdict
s6389251-chain49.gp	IMPOSSIBLE_SPAN ×1	Pass	Pass	5 L · 14 R (1–2%)	Fail
ALLPARTS-PLAYABLE.gp	clean	Pass	Fail	528 L · 837 R (64% · 100%)	Review
s6389251-BEST.gp = iter48	IMPOSSIBLE_SPAN ×1	Pass	Fail	blanket refret	Fail
ALLPARTS-SOURCE-FAITHFUL.gp	IMPOSSIBLE_SPAN ×2	Pass	Pass	—	Fail
c-tied.gp stem source	SPAN ×19 · TIE ×70	baseline	baseline	—	Fail
songsterr-ai-reference.gp	clean	n/a	n/a	—	Pass

Corrected: PLAYABLE's movement, measured against the minimal repair

A first pass here reported that PLAYABLE moved 100% of its healthy Rhythm notes. That number is raw position change and it overstates the fault. c-tied's Rhythm frets were written for capo 0 while the file declares capo 2, so any build keeping capo 2 must lower every fret by 2 just to be correct.

Build vs c-tied, Rhythm	At the minimal repair	Beyond it
ALLPARTS-PLAYABLE keeps capo 2	466 of 891 · 52.3%	425 of 891 · 47.7%
s6389251-chain49 drops capo to 0	842 unmoved · 94.5%	48 of 891 · 5.4%

PLAYABLE relocates 47.7% of the Rhythm staff beyond what the repair required. The over-relocations run to a higher string at a much lower fret: (+1 string, -7 fret) ×128, (+1, -11) ×31, (+2, -12) ×26. That is the cost function that likes low frets walking the part toward the nut, and it is the same shape as iter48 wearing a different filename. The corrected figure is smaller than the first one, and it still separates the two builds by an order of magnitude.

FindingThe Rhythm capo mutation, classified

c-tied carries capo 2 on the Rhythm staff and chain49 carries capo 0. The tuning is byte-identical in both. That change had to move either the frets or the sounding pitches, so it needed a verdict before anything downstream could be trusted.

c-tied.gp · Rhythm

capo2

tuning36 45 50 55 59 64

notes891

MIDI matches fingering0

MIDI differs891 · delta −2

Every note contradicts itself. The frets were written for an uncapoed neck while the capo field claims 2, so the fingering sounds a whole step above the pitch the file declares.

s6389251-chain49.gp · Rhythm

capo0

tuning36 45 50 55 59 64

notes891

MIDI matches fingering891

MIDI differs0

The notation and the fingering agree on all 891 notes. Removing the spurious capo flag reconciled the file with itself and left the frets alone.

Classification: supported correction. The Lead staff carries capo 0 in both files and agrees with its own fingering throughout, so the defect was confined to Rhythm. A player reading c-tied's Rhythm tab with a capo at the second fret would have played the whole part a whole step sharp against the record.

Why the pitch gate passed c-tied anyway.
pitch_census() resolves each note as p = m if m is not None else sounding_pitch(tun, capo, s, f).
The stored MIDI value wins whenever it exists, and c-tied stores a correct MIDI value on every note. The gate therefore compared two identical pitch multisets and reported gone none / new none, while the fingering a guitarist actually reads was wrong on all 891 notes. The audio audit reads the other field, and that is what exposed it.

Phase 2Audio accuracy, same denominators across all four

Each staff is scored against its own stem lane. A note is supported when its measured salience beats all three controls: wrong pitch by one semitone, wrong pitch by six, and wrong time. Notes below 130 Hz are dropped from the numerator and the denominator together, so the unjudgeable low register cannot inflate anyone's score.

Transcription	Lead supported	Rhythm supported	Combined	Rhythm owner	Staff → stem	Control sep.
chain49	391 / 597 · 65.5%	426 / 656 · 64.9%	817 / 1253 · 65.2%	71.8%	L 2.29 dB · R 2.93 dB	87.9% @ 10.5 dB
songsterr	118 / 215 · 54.9%	418 / 1018 · 41.1%	536 / 1233 · 43.5%	73.7%	L 2.88 dB · R 3.52 dB	74.0% @ 5.2 dB
c-tied stem source	367 / 572 · 64.2%	167 / 677 · 24.7%	534 / 1249 · 42.8%	45.6%	both → L · 0.87 dB	70.5% @ 4.8 dB
SYNTH control ceiling	236 / 281 · 84.0%	269 / 347 · 77.5%	505 / 628 · 80.4%	83.7%	L 21.8 dB · R 19.3 dB	99.1% @ 25.5 dB

The capo finding and the audio audit corroborate each other from independent directions. The audit scores sounding_pitch(tun, capo, s, f), which is what a player physically produces. c-tied's Rhythm staff scores 24.7% while chain49's scores 64.9% on the same recording, and the whole-step error is a complete mechanical explanation for that gap. Neither measurement was built to test the other.

On stem ownership. The critique's figures of 70.5% separation and 4.8 dB margin belong to the c-tied audit, where both staves map to the left stem and Rhythm holds a 0.87 dB margin. Ownership there is genuinely unresolved. chain49 separates its staves onto distinct lanes at 2.29 dB and 2.93 dB, and the synthetic control reaches 99.1% at 25.5 dB, which establishes that the detector behaves on cleanly separated material. Stem attribution still stays out of any accuracy claim.

BlockerThe single remaining hard failure, located

IMPOSSIBLE_SPAN — bar 62.938, Rhythm Guitar
written s0f11 s1f4 span 7 ties none
sounding [47, 49] attacked [47, 49]
held span 7 frets against a 4-fret hand at 128 bpm
simplest legal alternative → span 7: s0f11 s1f4

On this tuning the pitch 47 sits at s0f11 or s1f2, and 49 sits at s1f4 or s0f13. Every complete voicing of both notes together spans at least 7 frets, so no refingering closes this. The gate is correct to refuse it, and refingering is the wrong instrument for the job.

Superseded, and my read was wrong twice

This section is stale and is kept as a record of two errors. The iMac cleared this span in its Phase 3 before the bundle reached me, by removing Note id 2407 on the Rhythm staff at beat 247.750, taking span 7 to span 0 and moving FAIL to REVIEW.

1I called pitch 47 unmeasurable and unresolved. The iMac measured it at 35.88 dB in the Rhythm's own stem, passing all three controls: +6.1 against the semitone, +12.0 against the tritone, +5.8 against wrong time. The 130 Hz floor in the audit build I was handed excluded it. A later run did not.
2I withdrew the pitch 49 failure as ambient, and that withdrawal was the error. Pitch 49 fails wrong-time by 18 dB where the window's median failure is 2.53 dB. Eighteen is seven times ambient, so the signal was real and I talked myself out of it. The iMac read it correctly: pitch 49 is louder in the Lead's stem, the Lead writes it one beat either side at 246.750 and 248.250, and the chord was two guitars on one staff.

A weak-margin region lowers confidence in marginal calls. It does not license discarding a call that clears ambient by seven times.

Original reading, retained for the record

Both notes were looked up in audit_chain49.json at beat 247.75, where the gate's bar 62.938 lands at 128 bpm under bar = 1 + beat/4.

Note	f0	Measurable	Salience vs wrong-time control	Read
pitch 47 · s0f11	123.47 Hz	No under 130 Hz floor	not scored	Unresolved
pitch 49 · s1f4	138.59 Hz	Yes	sal 31.87 against ctrlT 49.90	Unresolved

Verdict: unresolved on both notes. Pitch 47 sits below the low_hz: 130.0 measurability floor, so it is excluded from both numerator and denominator and stays unresolved until the source mix arrives.

Pitch 49 was first read here as NOT SUPPORTED. That read was wrong and is withdrawn.
It loses to its wrong-time control, and so does the median note in that whole window. Across beats 224 to 256 the median sal - ctrlT margin is -2.53 dB, against +9 to +14 dB everywhere before beat 160.
A note failing where the median note also fails carries no information about that note. The measurement has no discriminating power there.

The actionable consequence: do not refinger this beat, and do not promote past it. The evidence at bar 62.938 is too weak to convict or acquit the chord, and the lower note needs the source mix regardless.

The regional dip, now explained

First hypothesis, refuted. Low register does not explain it. Spearman rho between the share of sub-130 Hz notes and the support rate is -0.236 across 13 windows, which is weak, and the windows at beats 288 to 352 carry 17 to 20 percent sub-130 Hz notes while scoring 68 to 74 percent support.

What it actually is: the guitar signal drops and the controls hold still.

Beats	Median sal	Median ctrlT	Margin	Support
0–160	40.2 – 45.1	28.5 – 31.7	+9.5 – +14.3	67 – 89%
160–256 the dip	31.0 – 32.1	29.3 – 34.6	+1.7 to −2.5	41 – 46%
288–384	52.4 – 53.7	27.8 – 31.7	+20.7 – +25.4	68 – 76%

The control median stays flat between 25.8 and 34.6 dB across the entire song. The signal median falls by roughly 10 dB across beats 160 to 256 and then rises by roughly 21 dB after beat 288. The margin, and with it the discriminating power, collapses in the middle.

It is a property of the recording rather than of any transcription. The dip appears in all three at the same beats.

Beats	chain49	c-tied	songsterr
128–160	73.2%	33.8%	50.0%
160–192	45.9%	4.9%	51.2%
192–224	40.7%	12.7%	36.5%
224–256	41.0%	25.0%	29.4%

No transcription is being penalised for a defect here. Beats 160 to 256 are a low-signal passage in the right guitar stem, and every candidate loses margin there together. Any verdict drawn from that window should be treated as low-confidence for all three.

FixedThe gate now checks both pitch fields

NOTATION_CONFLICT was added to tools/impossible_gate.py as a hard code. It compares every note's stored MIDI against sounding_pitch(tun, capo, s, f) and hard-fails on disagreement. The original is backed up alongside it.

Build, Rhythm staff	capo	Conflicts
c-tied.gp	2	891 of 891
s6389251-chain49.gp	0	0
ALLPARTS-PLAYABLE.gp	2	0
s6389251-iter48.gp and BEST	2	5

Two builds repair the same c-tied defect by opposite routes. chain49 changes the capo to match the frets. PLAYABLE changes every fret to match the capo. Both reach zero conflicts. iter48 and BEST attempted the PLAYABLE route and left 5 notes unrepaired, which nothing had reported before this gate existed.

Regression, run with PYTHONPATH set to tools/
_tests/test_comparison_basis.py — 8 of 8 correct
_tests/test_promotion_rule.py — 38 of 38 correct
No verdict changed anywhere except c-tied and iter48/BEST gaining the new code. songsterr-ai-reference.gp still PASSes, so the check does not fire falsely.

LedgerReceipts present on this machine

Eleven receipts ship in the bundle. Receipt 6f5b0815969eca47, named in the critique, appears in none of them, and it appears in nothing this machine ran. The three runs below marked local are mine.

7d1254e2 · 2d4e5917BEST / iter48 · preservation PASS, position FAIL

52942b12 · 4779832fALLPARTS-PLAYABLE · verdict REVIEW, position FAIL

5538dec2 · 13f93f5eALLPARTS-SOURCE-FAITHFUL · position PASS, IMPOSSIBLE_SPAN ×3

b58a3301span88438, the parent chain49 replaced · not shipped in the bundle

145164704462de17 · position PASS, SPAN ×1 + TIE ×1

05723529chain49 under the old gate 7a4696cd · position PASS

236e383a · bc363696chain49 under gate ece41376 · local, preservation and position NOT_RUN

Preservation and position read NOT_RUN on the two local receipts because those tiers key off --source and --checkpoint, and the parent build s6389251-span88438.gp was left out of the bundle. Supplying that file restores both tiers against the true parent.