msm-corpus-fit: MSM fitting pipeline, statistical Turing test, and committed corpus results

The complete msm-corpus-fit package: the Markov-switching-multifractal model zoo and exact-likelihood fitter, the DFA/MFDFA estimators, the parameter-recovery gate, the statistical Turing test, the competitive multifractal battery ported from Zhang, Sun and Xiong (2025), the crosswalk’s null calibration, the pinned corpus download and preprocessing scripts, the pytest suite, and every committed result under out/ — the fitted per-piece parameters for 1,264 MAESTRO piano performances and 198 Groove drum takes, the competitive records v1/v2/v3 with their addendum, the voice-set and envelope-null records, the null-floor records v1 and v2, and their figures.

Headline results the package reproduces. Real solo-piano onset density is strongly persistent and multifractal, inside the instrument’s own published band; the fitted model retains the persistence but does not pass the sufficiency test; drum-kit counts are near-white with the structure on the velocity channel instead; and, in the competitive arm, the project’s own asymmetric hypothesis inverted — every arm of the instrument sits outside the human DFA-α band, an excess that survives a common scale band and whose causal locus is per-note energy decay in the audio layer.

Reproducibility split — what runs from this deposit and what does not

Runs from this deposit alone (plus the pinned public downloads the scripts fetch and sha256-verify): the MAESTRO and Groove corpus fits, the statistical Turing test, the synthetic null panel and the null-floor curve, the parameter-recovery gate, and the whole test suite. The competitive arm re-scores Zhang, Sun and Xiong’s (2025) published per-song baseline CSVs, which the deposit fetches and verifies rather than redistributing.

Does not run from this deposit alone: the GiantMIDI arm of the null calibration. It reads a CC BY 4.0 derived onset series (note-attack timestamps) that is committed in a private sibling repository of the same program, resolved by a relative path; the arm is skip-if-absent, so the driver runs and simply omits that panel. That arm is the record’s only real-data decorrelation evidence, so this limitation is stated here rather than discovered. Everything else, including the corrected v2 floor curve and its regression tests, reproduces from pinned seeds.

Scope of the dispersion result (dated correction, 2026-08-30)

The record’s one positive claim — that the engine out-disperses Suno, DiffRhythm and YuE on alpha_peak (1.248× the Billboard IQR) and spectrum_skew (0.858×) — holds under the IQR ratio only. The same committed CSV computes a standard-deviation ratio and a span ratio, and under both of those Suno out-disperses the engine on both descriptors (and YuE does under the std ratio). The IQR is defensible on heavy-tail grounds (Billboard’s alpha_peak has skew +3.14, excess kurtosis +12.5 and std/IQR 2.55), but that rationale was written after the statistic was seen to win, so the dispersion statistic is being pre-registered before the pending full run. The Jensen-Shannon excess ordering carries two further caveats — the floor is drawn under a different sampling design from the statistic, and the statistic is bounded, so additive excesses across floors differing by 3.4× are not commensurable; the scale-free contrast excess/(1−floor) preserves and widens the ordering (engine 0.515 > DiffRhythm 0.311 > YuE 0.206 > Suno 0.090).

A null under the width family, in the envelope domain (E8)

An envelope-domain IAAFT surrogate null finds 5 of 12 shipped packs’ Δα indistinguishable from a linear surrogate — the first null ever placed under the width claim in the domain the competitive descriptors live in. Two disambiguations are mandatory: it measures Δα on audio envelopes, not the two-point MFDFA width on event-time onset streams that the catalogue-wide IAAFT result is measured on; and IAAFT truncates rather than pads, so every series is scored at 16,384 samples (109.2 s) and no number from it is comparable to the human corpus band. Its per-descriptor table runs 168 dependent rank tests at a p floor of 0.05 with no multiplicity control, so counts near 1–3 of 12 are not separable from chance.

The null-floor calibration was re-cut (v2, 2026-08-30)

The crosswalk’s null calibration is deposited at v2. The v1 record generated its iid “Fano-matched” null streams at the fitted latent Cox rate where the generators take a marginal mean, so its floor curve was measured at 0.577× the corpus’s own operating point. v2 re-runs the calibration mean-parameterized, adds the paired per-piece decorrelation tests, the unclamped surviving fraction, the off-crosswalk-domain counts and a tempo-grid sensitivity arm. v1 is retained unedited, with a dated pointer, because sealed records in this program are append-only.

Data provenance and attribution

This record uses the MAESTRO v3.0.0 solo-piano corpus (Hawthorne, Stasyuk, Roberts, Simon, Huang, Dieleman, Elsen, Engel and Eck, “Enabling Factorized Piano Music Modeling and Generation with the MAESTRO Dataset”, ICLR 2019; hosted by Google Magenta) and the Groove MIDI Dataset v1.0.0 (Gillick, Roberts, Engel, Eck and Bamman, “Learning to Groove with Inverse Sequence Transformations”, ICML 2019).

The raw corpora are not redistributed here. They are fetched at run time from their public sources by pinned download scripts that verify each archive against a recorded sha256 (MAESTRO: maestro-v3.0.0-midi.zip, sha256 70470ee253295c8d2c71e6d9d4a815189e35c89624b76d22fce5a019d5dde12c). Only derived quantities — fitted per-piece parameters, DFA/MFDFA statistics, model-selection verdicts, and the figures built from them — are deposited.

MAESTRO is licensed CC BY-NC-SA 4.0 (non-commercial, share-alike), which controls over this record’s CC BY 4.0 grant wherever the two conflict, for any MAESTRO-derived material. The Groove MIDI Dataset is CC BY 4.0, so Groove-derived outputs carry no additional restriction.

Archive built from the repository at commit f3970fbe; gitignored data and caches are excluded by construction.

Part of the Feno program. The source repository is private; the reproduction materials for this record are contained in the record itself.