Shiner · BELIEVEYOUME, Tab Workbench

All ten songs have a Songsterr tab, a .gp file, and a stem folder. What separates them is whether the stems can actually prove the guitar part. Two songs were separated to four stems only, so the guitar is buried inside other and no per-attack guitar evidence exists for them.

ALL 10Songs 9 and 10, and the audit is complete

Asleep in the Trunk: 88.9% attack recall on a file that is almost empty

Jackie: the cleanest map on the record

Jackie's map beats The Mutiny's on every health measure: 1.7% clamped against 6.0%, a max collapse run of 12 against 66, and 0.984 unique audio times per attack against 0.940. Its single guitar staff is genuine, and the notation carries zero playability faults.

PLANFinishing every /goal requirement

The /goal condition asks for two things in order: audio accuracy against the stems reported as measured numbers, then physical playability. Below is every requirement in that condition, against what The Mutiny has been measured to hold.

Phase 1. Close the three executable gaps, no new research needed

1. Implement duration and tie agreement. written_attacks() already returns duration_beats per note, so the data is there and the comparison against decay-to-half-energy is missing. Ship it with a shifted-time control like every other measure, and report it refused if it loses.
2. Itemise the 795 unmatched attacks. The earlier attempt saturated at 100% because it rebuilt the map with np.interp. The tool's own BM.apply_map(tab_t, rec_t, times) is the call that produces 74.1%, so extraction goes through that call and yields each miss with its staff and bar.
3. Classify every miss as supported correction, unsupported change, or unresolved ambiguity, which is the wording that /goal uses. On present evidence most will fall under unresolved, since an onset detector misses attacks inside a held chord, and saying so per note is the requirement.
4. Initialise the checkpoint ledger for The Mutiny with --init-checkpoint so any future candidate has a baseline to beat.

Phase 1 executed, all three gaps closed

Run 2026-08-17 through the audit's own BM.build, BM.apply_map, nearest and DECOY_S, so the numbers match what that tool reports rather than a reimplementation. New tool: miss_and_duration_audit.py.

Gap 1. Duration and tie agreement, now implemented

The /goal line item is satisfied: the measure exists, ran, and is reported with its control. It joins the refused list rather than the usable one.

Gap 2. The 991 misses, itemised and classified

Song-wide the attack measure discriminates hard: 74.1% real against 11.1% for the decoy at +7.3 s, a 6.7x separation. Per bar it never isolates a discrepancy. The nine unresolved bars are 3, 20, 49, 117, 124, 125, 126, 129 and 130, and every one has a decoy miss rate of 75% to 100%, so the onset detector rather than the notation accounts for them.

Gap 3. Checkpoint ledger

Initialised for The Mutiny. CHECKPOINT.json and RECEIPTS.jsonl now sit beside the song, status best_validated_so_far, so any future candidate has a baseline to beat rather than a filename to argue from.

Song 2 of 10: The Alligator, full pipeline

Song 3 of 10: So Far So, blocked at the map stage

All three stems fail on all four criteria, so no per-note audio metric is interpretable for So Far So. The cause is onset yield: 83 to 132 onsets across 259 seconds is 0.32 to 0.51 per second, against The Mutiny's 2.86. Song-level alignment and attack recall stay reportable; nothing per-note does.

Both playability faults involve ties, the same shape that turned out to be a held pedal on The Mutiny:

The Lead staff also carries the same TUNING_EVIDENCE review as The Mutiny: string 0 detuned by 2 with its open pitch never used in that staff.

Song 4 of 10: My Mirror Hates Me, the cleanest so far

Four songs profiled, side by side

Song 5 of 10: Endless Summer, and the session's first successful correction

I then suspected the tooling mishandled the meter, since ASE.events derives each bar's length from its notes and defaults to 4.0 for an empty bar rather than reading the time signature. Measured, that hypothesis is refuted: drift is -0.4 s across 248 s on Endless Summer and 0.0 s on The Mutiny, because filled bars derive their span correctly.

The two shapes, and why they were safe to repair

Every one of the 20 records carries ties: [], so the tie-chain trap that regressed The Mutiny candidate does not apply. Both alternatives were verified pitch-identical before any edit.

The validated candidate

Song 6 of 10: Not Too Much, 15 spans down to 1

The health gate caught a high-gain degenerate map for the third time

The Rhythm stem scores the highest gain of the three on 3 anchors and 60 onsets, clamping 58.3% of the tab and collapsing 568 consecutive attacks onto one audio time. The stem that actually carries the tab is Other Guitar, which scores lower and is the only one that is interpretable.

14 of the 15 spans were repairable, and 1 was not

The two repaired shapes share the same three pitch-identical moves in D Standard: s1f5 to s2f0, s2f7 to s3f2, s3f7 to s4f3. The third shape holds ties on three strings and its own suggested alternative is also span 7, which is the held-chord pattern that proved to be correct notation on The Mutiny, so it was left alone. The repair script skipped those two beats by itself and said so.

Both candidates, with receipts

Song 7 of 10: Lazarus, blocked on both maps, with forced octave leaps

The Phase 0 re-pull could not be completed as a .gp: Songsterr serves notation as JSON per track and offers no .gp export over plain HTTP, and the only other Lazarus file on disk, _gp_extract_Shiner-Lazarus-05-04-2026.gp, holds the identical 138 bars and identical three tracks. So the audit ran on the local file, and the live s5476396 with its vocal and drum staves remains the better source that nothing local can yet supply.

Both maps fail, so no audio verification is possible

Thirteen onsets across 281 seconds is 0.05 per second. Lazarus joins So Far So as a song where no stem yields an interpretable map.

The octave leaps are forced by the fretboard

136 SHIFT and JUMP records concentrate on one string: s5 carries 210 of them, s4 carries 60, s2 carries 2. The shape is the same throughout, an octave leap on a single string:

Pitch 82 has exactly one placement on a Drop C guitar, s5f20. Pitch 70 has four: s2f22, s3f17, s4f13, s5f8. So the high note cannot move, and the only refret available is to raise the low note to s4f13, which shortens the leap from 12 frets to 7. The gate flags a 7-fret move in 128 ms as a JUMP too, at 55 frets/sec, so that repair trades one flag for another.

Correction: the drive being absent blocks LESS than I said

I reported the work as blocked on the missing T7 Shield volume. That was too broad. A complete copy of all ten .gp files sits on the internal disk at ~/Desktop/Brandon-Vault/New Folder With Items 2/New Folder With Items 7/IF YOU CAN'T FIND SOMETHING ITS HERE/BandBooks/BELIEVEYOUME/_canonical_2026-06-04/songs/. What is genuinely missing is the STEMS, which exist nowhere on the internal disk apart from a drum-only extract.

Song 8 of 10: Broken Satellites, the notation half

Songs 9 and 10: Asleep in the Trunk and Jackie, notation half

Asleep confirms on the local copy exactly what Phase 0 found against the live tab: our file is a shell while s5085156 holds 7245 notes. Jackie's single guitar staff is genuine rather than a local defect, since the live tab also returns one.

The field-standard metric, finally run, and it is weak

mir_eval.transcription note-level F-measure is the MIREX standard and the one thing this pipeline had never executed. It pairs onset AND pitch into a single matched-note decision, which is neither of the questions the earlier measures asked. It has now been run. It needed no drive: the tabs, the full-mix mp3s and per-instrument YIN transcriptions of the stems are all on the internal disk.

Worst-case discrimination is 1.28x at 50 ms and 1.16x at 250 ms, both under the 1.5 bar. Endless Summer is the closest control on both note count and duration, and it is the one the metric struggles hardest to separate. Attack recall remains the leader at 5.70x.

mir_eval with a POLYPHONIC reference: it works, and no install was needed

The monophonic YIN run gave a weak 1.28x and was not a fair test of the metric. The fair test needs a polyphonic reference, and basic_pitch crashed with SIGFPE on this box. The fix turned out to be free. Its default model is nmp.mlpackage, CoreML, which is what crashes. nmp.onnx ships in the same directory and onnxruntime 1.23.2 is already installed, so selecting the ONNX backend fixed it. No tensorflow, no tflite-runtime, no download.

Broken Satellites, audio half, and it needed no drive either

The local full-mix mp3s cover all ten songs, so the audio pipeline runs on them exactly as it ran on The Mutiny. The isolated stem would be better and is not required.

Both supported bars carry 40 and 90 attacks with 100% miss against decoys at 80.0% and 83.3%, so they clear the threshold by a wider margin than The Alligator's did. They are listening candidates rather than proven defects, and the duplicate-staff inflation applies to them too.

Phase 2. The decision that is not mine to make

After Phase 1, /goal on The Mutiny reduces to one question. Nine measures were tested with controls on both sides and only attack recall survived, so today a corrected file cannot be justified note by note. Three ways forward:

• Accept the inherited tab with its documented profile: 74.1% attack recall, 88.0% precision, 48.2 ms median timing, zero genuine playability faults. Ship it, and write the standard down.
• Fund the missing-signal work. The search space is mapped: map-based continuous features fail, map-free discrete measures fail, chroma has no power. What is untried is a learned or template-based matcher rather than a hand-designed statistic.
• Correct by ear with the tools as evidence rather than as the licenser, which is the route that your own standing rule already prefers when a scorecard and your ear disagree.

Phase 3. The other nine songs, with the real cost

• Ready now, five songs: Not Too Much, The Alligator, So Far So, My Mirror Hates Me, and Endless Summer. Each needs a map build, a health gate, the audit, the playability tier, and the triage.
• Needs a re-pull first, two songs: Lazarus and Broken Satellites, both settled in Phase 0 and neither pulled yet.
• Blocked on stems, two songs: Asleep in the Trunk and Jackie hold four stems each with the guitar inside other, so they need six-stem re-separation before any guitar question.
• One song is the pilot already run: The Mutiny.

00Phase 0 result: run, measured, closed

Executed 2026-08-17. Live notation pulled per track off the Songsterr CDN route st_gpdiff.py uses, then counted bar by bar. Four of the five Phase 0 items are now settled by measurement rather than by inference.

Per-track detail, live

00bPhase 1 result: The Mutiny, measured

Run 2026-08-17 against 08_the_mutiny/songsterr.gp, sha256 21b37b73b6f35faa, 135 bars, 4906 notes. Every number below is validator output rather than a claim.

Step 1. Two guitars, or one part split across two staves

Step 2. The beat-to-seconds map

Three stems were tried at offset 4.04s, 130 bpm. The right channel wins by a wide margin, and the reason matters: the stereo sum masks transients that the isolated channel keeps.

The accepted map covers 126 of 135 bars, mean bar length 1.8469s against a nominal 1.8462s. The last nine bars fall beyond the final anchor and are not measured. Passing the chance test alone was not enough: the first bar-length reading looked catastrophic until the extrapolation past the anchor span was removed, so the covered span is stated explicitly rather than implied.

Step 3. Audio accuracy against the right-channel stem

Step 5. The register-aware fix, tested and refuted

The guitar stem holds under 2% of its energy below 150 Hz, so the obvious repair is to read the low band from the full mix. The mix at Music to Stem/Beleiveyoume/BELIEVEYOUME (2025)/03 - The Mutiny.flac runs 251.490295s against the stems' 251.490295s, identical to the sample, so no time correction was needed. A combined file was built with a 4-pole crossover: guitar right channel above 150 Hz, full mix below it.

Then a negative control, which is the test that settles it: the same Mutiny tab measured against So Far So's guitar stem, a deliberately wrong pairing.

Step 6. A correction protocol on the validated metric, built and refused

With pitch, ownership and attack-string all disqualified, the only surviving audio metric is attack support. A new tool, bar_support_audit.py, localises it: rather than judging a note, it asks which BARS have no acoustic support while their neighbours do. A single unmatched attack proves nothing, since an onset detector systematically misses attacks inside a held distorted chord, and a region carries information a note cannot.

It ships with a decoy control, the same tab shifted 3.7 seconds. Two constructions were tried and both were refused:

Cause: 718 onsets across 251 seconds is one every 0.35 seconds. The DTW map compresses any input into that distribution, so a tab shifted by nearly four seconds still falls beside an onset. Onset density this high leaves the measure no headroom, which is a property of dense distorted rhythm guitar rather than a coding error. The tool is kept, exits nonzero, and emits nothing whenever the decoy keeps up.

Step 7. Replication on a second song, which refutes the material explanation

The obvious reading of Step 5 was that the material defeats pitch matching, since dense distorted power chords in Drop C are close to chroma-stationary. That is testable. My Mirror Hates Me differs on every property that could explain it: Standard E rather than Drop C, 114 bpm rather than 130, and 1.59 onsets per second rather than 2.86.

Attack recall stayed correctly ordered on both songs, 74.1% against 13.0% on The Mutiny and 66.3% against 44.6% here. Its discrimination is much weaker on this pairing, 1.49x against 5.70x, because the two songs run 287s and 251s and the warp absorbs some of the difference. Attack recall remains the only metric that has never pointed the wrong way.

Step 8. The mechanism, found, and the gate that catches it

Step 7 said the defect was in the measure's construction, which was as far as the evidence went at the time. It goes further. The cause is a degenerate DTW map, and it is measurable directly.

A wrong pairing yields few anchors, so interpolation pins a quarter of the tab outside the anchor span and collapses up to 284 consecutive attacks onto a single audio time. Every one of those notes is then compared against the same frame. When that frame is a sustained chord, a large share of written pitches agree with it, while the shifted-time control falls on a different and sparser frame and collapses. That is the large spurious separation, appearing on the pairing that is most wrong.

This changes the standing conclusion. Pitch agreement is not inherently inverted. It was being fed maps that could not carry it, on pairings nothing was screening. Whether it becomes usable on a map that passes the health gate is the next thing to measure, and it now has a gate in front of it.

Step 9. Pitch on healthy maps, measured three times

With the health gate in place the earlier comparison can be redone properly. The wrong pairing that appeared to invert the metric fails the health gate on all four criteria, so it is disqualified as a control. What remains is pitch measured only through maps that pass.

The [L] and [R] comparison is the sharpest form of it. Two channels of the same song at the same instant, both maps healthy, and pitch scores 53.4% against 52.3%. Attack recall separates them 2.1 times over, 74.1% against 34.9%, so the tab's attacks clearly follow the right channel. Chroma cannot see a difference that attack timing makes obvious.

Step 10. The feature survey, all four candidates refuted

The natural next question is which audio feature can order a correct pairing above an incorrect one when chroma cannot. The obvious control, My Mirror's tab against The Mutiny's stem, fails the health gate on all four criteria, so it is disqualified before any feature is computed. The one healthy-map pairing that is nonetheless wrong is The Mutiny's own left channel: both channel maps pass, both cover the same song at the same instant, and attack recall already orders them 2.13 to one.

Four candidates, each sampled at the tab's 3068 attack times mapped through that channel's own healthy map, each against a within-channel control at the same times shifted 3.7 seconds.

The searchable space narrows to discrete event matching rather than continuous spectral features. Onset strength, spectral flux, transient clarity and attack envelope sharpness are all ruled out on this material with proper controls, so a future attempt should start from event-level questions rather than from frame-level ones.

Step 11. The one hard fault is a false positive, and the tab is right

Playability needs no audio, so the audio blocker never applied to it. Dumping the tie state across bars 3 to 8 resolves the single hard fault on the file.

String 3 fret 5 is a pedal held by an unbroken tie chain from bar 3 through bar 8 while the melody moves on string 4 through frets 8, 9 and 10. One finger holds, another reaches. That is index and pinky across adjacent strings at position 5, which every guitarist does. The gate models it as four fingers fitting inside four frets at once, which is the wrong model for a two-finger pedal and reach.

Step 12. Discrete measures without a map, also refuted

Step 10 concluded the surviving signal class was discrete event matching rather than continuous features. That is testable without a DTW map at all, which sidesteps both the circularity of Step 6 and the map degeneracy of Step 8. Tab attacks placed at linear time from bpm alone, compared against the audio's onset stream as a sequence.

Step 4. Playability, which needs no audio

The single hard fault, in full:

The Lead REVIEW says string 0 is detuned by 2 from the rest of the tuning and its open pitch is never used in that staff. The whole guitar is in Drop C for the Rhythm part, which uses the low C. Read on its own, the Lead staff gives no evidence for the drop. That is a review flag rather than a fault.

01Vocal harmony: the direct answer

Harmony does exist, written as stacked notes on that single staff. Five of the ten stack notes inside the vocal part, up to three at once. The other four are strictly one note at a time, so any harmony those songs actually sing is absent from the notation entirely.

02Note content, every staff

Counted out of the GPIF by walking MasterBar to Bar to Voice to Beat to Notes. This is the measure that separates a real transcription from a correctly-shaped empty one.

03The fix plan, in order

Ordered by dependency. Nothing below Phase 0 is worth doing against a file that has not been re-pulled and content-checked, because four of the issues found so far turned out to be defects in the local copy rather than in the source.

Phase 0. Trust the source before touching anything

Phase 1. Calibrate the guitar split on the one song with ground truth

Phase 2. Bass

Phase 3. Vocals, which is the harmony question

Phase 4. Unblock the two songs with no guitar evidence

Phase 5. Put the corrections back

04The verdict, in one table

05Songsterr registry

Every ID resolved with a following curl -sIL on 2026-08-17. Tempo, tuning, and time signature are the values Songsterr displays; they were read off the site in the 2026-06-03 audit and match the tuning bytes inside each .gp, which I re-read today.

06Guitar Pro track census

Read straight out of each Content/score.gpif today. Tuning is given as MIDI pitch numbers low to high, exactly as the file stores them.

07Stem census

Two roots hold BELIEVEYOUME stems. Moises_Stems/BelieveYouMeStems/ is the superset with all ten songs; Full wav files for music books are here/ holds the same eight and is missing Asleep and Jackie. Eight of the ten stems_pointer.txt files point at the smaller root.

08What Songsterr holds that the local export does not

The canonical folder was exported 2026-06-02/03. Songsterr has moved since. Every row below was read from songsterr.com/api/meta/<id>/revisions and the matching revision endpoint on 2026-08-17, so the track titles and tuning come from the live revision rather than from the local file.

09Four traps that will bite

10The correction queue

1. The Mutiny: start here. It is the only song carrying real RhythmGtr [L] and [R] files, so the two-guitar split can be validated against ground truth rather than inferred. Whatever the splitter concludes on Mutiny is the calibration for the other nine.
2. The Alligator: second, because it has an audio_truth.gp3 from the 2026-06-02 stem pass. That gives a prior to regress against, and a resolved key at high confidence.
3. Not Too Much, So Far So, My Mirror Hates Me, Endless Summer: the clean middle. Full guitar layers, two GP staves, vocal and bass stems present.
4. Lazarus: re-pull s5476396 first. The vocal and drum staves exist there, so nothing has to be built from scratch.
5. Broken Satellites: re-pull the live tab, decide between s5084035 and the 2026-07-08 s5965143, then correct.
6. Asleep in the Trunk, Jackie: re-pull the 2026-07 hand-built revisions, re-separate the audio to six stems, settle the Asleep time signature, then queue. The stem gap is a disk fact and does not change with the tab.

11Canonical paths

• Tabs, GP, charts, lyrics: /Volumes/T7 Shield/BandBooks/BELIEVEYOUME/_canonical_2026-06-04/songs/<NN_slug>/, each holding songsterr.gp, songsterr.mid, fromcdn.mid, 3guitar.mid, chart_GBV_TAB.md, lyrics.txt, stems_pointer.txt
• Stems, all ten songs: /Volumes/T7 Shield/Moises_Stems/BelieveYouMeStems/
• Stems, eight songs, redundant: /Volumes/T7 Shield/Full wav files for music books are here/
• Resolved keys and method trails: /Volumes/T7 Shield/BandBooks/BELIEVEYOUME/keys_resolved.yaml
• Songsterr metadata audit: /Volumes/T7 Shield/BandBooks/BELIEVEYOUME/SONGSTERR_AUDIT_2026-06-03.md
• Stem versus MIDI disputes: /Volumes/T7 Shield/BandBooks/BELIEVEYOUME/STEM_VS_MIDI_AUDIT_2026-06-03.md
• Three-guitar derivation work: /Volumes/T7 Shield/BandBooks/BELIEVEYOUME/3-guitar-derivation/

Skills this feeds

• /impossible-guitar-parts: grades a staff for playability and finds the two-guitars-on-one-staff cause
• /composite-stem-alignment: the beat-to-seconds map the audio checks depend on
• /songsterr-tab-guide and /songsterr-upload: build, diff, and push corrected tabs back, preflight first
• /five-stem-song-analyst and /audio-stems-to-midi: the vocal and bass evidence layer