Two guitar split

Which songs carry a recoverable second guitarist, measured from the stereo guitar stem. 26 songs, 2026-09-08.

What the number means

Margin is the whole verdict. It is the share of one channel's attacks that land on an attack in the other channel, minus the identical test with that channel pushed +5 seconds. A raw coverage percentage with no control is what produced the earlier “not separable” reading on a song that separates at 79.6.

The corpus

Why the existing tabs cannot be split note by note

The obvious move is to take the tab we already have and hand each written note to whichever channel played it. That was tried and it does not survive its own control. Every position on Kilgore Trout's 449-position rhythm staff was scored, with the offset searched over ±3 seconds at 10 ms steps:

Loosening the window raises the real score and the control together. A 90% match at 200 ms is density, not alignment. There is no tolerance at which this tab's attack times can say who played what, so the tool refuses instead of writing a confident two-staff file out of noise.

python3 ~/.claude/skills/two-guitar-split/split_stem_channels.py \
    --stem "<song>/guitar.mp3" --out-dir <work>/split

The gate runs first and refuses anything that is not SEPARABLE. Verified both ways: Kilgore Trout writes its two channels, Adriatic to Black Sea is refused as mono.

First song through the whole chain

The Immortal Soul of Mundo Cani, margin 70.8. Stem split, each channel transcribed on its own, both written onto two staves against the existing tab as the template.

For scale, the inherited AI tab carried 128 and 237 guitar note positions. This one carries 529 and 633, and every one of them came from the recording.

The audio accuracy audit, and what it can actually judge

Playability is the second question. The first is whether the notes are right, so the artifact went through audio_accuracy_audit.py and through a per-staff channel-ownership measure with wrong-song stems as controls.

Channel ownership, measured directly, against a human tab of the same song

Chance-normalised, because the wrong-stem controls differ wildly in onset density (1012 against this song's 96). The normalisation is validated by the +5 s shift control landing at −0.1 and +0.3 gain, which is zero.

Pitch evidence that alignment cannot break

Every alignment-based pitch measure on this catalogue fails its own control, because a tab whose attacks sit 105 ms off cannot be scored note against note. Pitch-CLASS content needs no alignment at all. What a staff plays across the whole song either matches what that channel plays across the whole song, or it does not. Reported as cosine similarity between the staff's pitch-class histogram and the channel's mean chroma.

The same measure on the human Songsterr tab of this song

Provenance

Zero note ids and zero beat ids in either guitar staff exist in the Songsterr AI tab. Of 529 and 633 written events, 25 and 13 happen to coincide with the AI tab on bar, position, fret, string and pitch, which is 4.7% and 2.1%, and they share no elements. No note content is inherited.

Timing, measured against the human tab

Timing was the open question, so it was diagnosed rather than restated. Every written attack was paired with the nearest onset in the stem, and the residual regressed against position to separate a wrong tempo from a coarse grid.

It is scatter, not drift. The slope is +0.0002 s per second at R² 0.002, and a full 5% tempo search buys 6 ms. So the tempo is right and no re-tempo can fix this.

Ties and note lengths, and the axis they unlocked

The first build wrote every note as one 16th slot, so it asserted no duration at all and duration agreement had nothing to compare. The writer now carries the audio-derived length of each note and ties it across the note values needed to express it. The result is the promoted artifact, s6285808-TWOGTR-WITH-TIES.gp, hash a5a4847130f540f9cf54ac782f79b826.

Every axis the validator scores

Chord contiguity and pseudo-fingerpicking, answered with the audio

The validator raises two findings that ask a question rather than assert a fault. INTERIOR_GAP is the chord-contiguity check and says "N attacked notes with M unexplained interior string(s); needs positive audio evidence". OWNERSHIP_SUSPECT is the pseudo-fingerpicking check and says "outer-string pseudo-chord across strings A and B; test whether one note belongs to the other guitar". Both were tested at their own instant against the two channels, with the same lookup at +5 s as the control.

Channel ownership per note

The canonical tool cannot do this, so it was measured directly: every attacked note on each staff matched by pitch and onset against BOTH channel transcriptions, within 60 ms.

The per-note breakdown, which is the part that answers ownership

Song 2: Secret, s6388915

Same chain, no changes. Artifact s6388915-TWOGTR-WITH-TIES.gp, hash 26a826fe6f686139. 95 bars, standard E, 75 bpm, 274 s.

Standing rule

Two guitars, never four. A third staff is never the answer, and the split is decided by who played the note, never by panning or by stem loudness.