Large language models can write music notation, but we lack a rigorous way to say whether what they write is structurally sound — and to steer them when it is not. We present Libretto, a validated, deterministic environment for symbolic music in which an LLM agent composes within a text grammar and is assessed and steered by a structural-assessment engine. The engine places any piece in a 29-axis fingerprint of token-derived, decorrelated structural features (rhythm, harmony, melody, texture, form), each scored as an empirical percentile against a frozen 314-song corpus; the axes are descriptive, not a learned “quality” score. We show the gates discriminate: real human music sits at a mean of 3.4 extreme axes (admitted), while degenerate generations pile up to 4–6, and a strict note-level copy gate cleanly separates novel material (≤0.29) from near-duplicates (≥0.31). On this substrate we add two interventions — a monotone-safe self-evolving loop that returns non-prescriptive “dosage” feedback to the agent, and mandatory retrieval of corpus-grounded idiom concepts plus centroid-nearest real exemplars. Across four applications (held-out gap-filling, component morphing, from-scratch generation, single-voice pedagogy), retrieval triples from-scratch pass rate (2/8→6/8) by de-degenerating the fingerprint, and the loop roughly triples gap-filling pass rate (12%→39%) and lifts from-scratch generation to 94%. Because every axis is deterministic and grounded, a failure is fully legible: a generated gap-fill’s “~2-chords-per-bar conveyor belt” and all-stepwise melody are pinpointed to specific axes at the 95th–100th percentile. The grammar round-trips to MIDI with 0.019% note loss, and the validated core is frozen under a sha256 manifest.
An LLM composes in the grammar; the deterministic engine scores it on the 29 axes and steers it — retrieve, compose, measure, and loop until it clears the gates.
One compact, human-legible text grammar — exact pitch, rhythm, and voices, and nothing else. The same text round-trips to MIDI; the LLM reads and writes only this.
KEY: C major | METER: 4/4 | TEMPO: 120 | GRID: 16th | BARS: 2 VOICES: Melody, Piano, Bass [C] Melody: E4@1>2 G4@5>2 C5@9>4 G4@13>2 Piano : C4+E4+G4@1>8 C4+E4+G4@9>8 Bass : C2@1>4 G2@9>4 [G] Melody: D4@1>2 G4@5>2 B4@9>4 D5@13>2 Piano : G3+B3+D4@1>8 G3+B3+D4@9>8 Bass : G2@1>4 D3@9>4
Note content is exact and reversible; velocity, micro-timing, and timbre are intentionally dropped.
Edit the grammar on the left — it renders to a piano-roll instantly, and you can play it. Runs entirely in your browser (no server).
↑ these structural axes are read directly from the text — every note's bar and slot is explicit, so they're exact lookups (off-beat notes are outlined). In ABC or raw MIDI you'd first have to reconstruct absolute timing (relative durations / tick deltas) before any of these is even well-defined.
Any piece is placed in a coordinate system of 29 token-derived structural axes, each a percentile against a frozen 314-song corpus. The axes are decorrelated (max |r| = 0.77, 12 of 406 pairs ≥ 0.5) and carry real — if soft — genre signal.
Percentiles are scored against a frozen reference of 314 real songs that we curate
across 8 genres from the Clean MIDI subset (~17k files,
2,199 artists) of the Lakh MIDI Dataset (Raffel, 2016; CC-BY 4.0). The 314-song selection and genre
labels are ours; the source MIDIs — and their artist/title, “with some inaccuracy”
— come from clean_midi (community transcriptions, hence anonymized Part{n} voices and
occasional near-duplicates, which the copy gate accounts for).
| genre | n | genre | n |
|---|---|---|---|
| Pop/Rock | 70 | Funk/Soul | 30 |
| Electronic | 34 | Jazz | 28 |
| Classical | 30 | Latin/Reggae/World | 28 |
| Folk/Country | 25 | Film/Score | 10 |
The gates are calibrated against ground truth and discriminate real music from degenerate or copied output — a deterministic, auditable target rather than a learned score.
| gate | real / good output | degenerate / copy | threshold |
|---|---|---|---|
| C1 non-degeneracy (#extreme axes/29) | real songs mean 3.4 | gap-task FAIL 4.36 vs PASS 2.25; un-grounded newgen 5.9 | budget 4–6 |
| Novelty / copy-risk (note-level — one metric, per-task reference) | gap-fill & newgen passing ≤0.29; education drills 0.065 | gap-fill flagged ≥0.31 (17/51 caught) | <0.30 vs corpus/answer · education <0.50 vs shown example |
| fidelity (grammar↔MIDI) | pitch exact, 0.019% loss | velocity/micro-timing/drums dropped by design | — |
Frozen core certified — distribution + encoder/decoder + metric definitions match a pinned sha256 manifest (v2.0.0); guard tests for the 29-axis shape, exact round-trip, determinism, and copy self-match all pass.
Two interventions sit on the substrate: mandatory retrieval (corpus idiom concepts + centroid-nearest real exemplars) and a monotone-safe self-evolving loop (≤3 rounds, pick-best) that feeds non-prescriptive dosage feedback back to the agent.
One grammar and one assessment engine support held-out gap-filling (n=51), component morphing (n=21), from-scratch generation (n=8/16), and single-voice pedagogy (n=32).
Audio rendered from the generated grammar (grammar → MIDI → General-MIDI synthesis). A few typical samples per task; the rest are browsable below.
One typical sample per genre. In the full list below, s1 / s2 are independent samples of that genre — separate generations that use the same prompt and the same retrieval, so they may sound similar (same style neighborhood; measured pairwise note-overlap ≈ 0.15, well under the 0.30 copy gate, so they are still distinct pieces).
Each clip glides from one style to another (real source → generated transition → real target).
A/B the model's generated region against the true held-out answer for the same gap.
Eight guided drills across keys & modes — listen, or download the MIDI and the engraved score (PDF).
Because every axis is deterministic and grounded, the metric’s verdict is not a black box. A failing gap-fill is legible as visible musical structure, pinned to specific axes.
Grounded diagnosis
FAILS with 6 extreme axes vs a budget of 3:
100th pct99th96th95th100th3rdNot a copy (copy-risk 0.134), rhythmically aligned (beat 86%) — the failure is purely structural, and every flagged axis is visible in the roll.
The package onboarding (libretto/ONBOARDING.md) — structure, what each task
implements, and how to reproduce. Scroll within the panel.
The Libretto package (grammar codec, the 29-axis engine, the loop and retrieval, the four applications) and the full results bundle are open.
@misc{xu2026libretto,
title = {Libretto: a deterministic structural-assessment
environment for symbolic music},
author = {Xu, Yichen},
year = {2026},
howpublished = {\url{https://github.com/Xyc-arch/Libretto}},
note = {Preprint; code and data available}
}