Musical Intelligence Reproduction Archive — Results Layer (1,097 verdicts, engine-free)
收藏资源简介：
This release supersedes the prior narrow Cycle 17 ds003720 artifact (v1.0.0)with the comprehensive Musical Intelligence (MI) Reproduction Archive. The archive reproduces 1,097 pass/fail verdicts (180 claim-style CSV records+ 917 pytest sub-tests) covering every numerical claim in the MusicalIntelligence paper. Reproduction runs without access to the MI engine sourcecode and without raw audio (reviewer cache-mode), completing in ~55 minuteson a MacBook Air M2 8 GB. Coverage spans the full MI three-layer architecture: • R³ perceptual front-end (01.1): 531 pytest tests; 13-corpus consonance battery; zero-calibration audit on 16,248 engine constants. • T³ temporal layer (02.1): 207 pytest tests on the 32-horizon, 24-morph demand-driven temporal grammar. • C³ cognitive core (03.1–03.7): 89 mechanisms × 131 beliefs F1–F8; ECE = 0.084 belief calibration; Cheung 2019 reward replication (β = −0.158, ΔAIC = −33.5, held-out r = +0.615); ChillsDB, TenseMusic, PMEmo, Eerola GEMS behavioural benchmarks. • C³ biological substrate (04.1–04.2): 132/132 neurochemistry accumulation + 11/11 pharmacological cross-validations (Putkinen 2025 μ-opioid PET 7/7, Salimpoor 2011 caudate-leads-NAcc 52/56 at +0.9 s); RAM topology (28/31 ≤10 mm hub-coordinate match, both nulls p<0.0001). • fMRI brain grounding (05.1–05.7): Mendelssohn Op.54 single-subject pilot; mech×region encoding on ds002725 (16/22 BH-FDR pass); voxelwise routing ablation on ds003720 (MI 4/4 vs MI-naive 1/4 vs Random 0/4 vs MERT 4/4 vs CLAP 2/4; CKA 0.994/0.414/0.125); cross-paradigm fingerprint ρ = +0.998. • Portfolio falsifiability (06.1, 06.3): five pre-committed Table 5 cells all PASS; from-scratch AI baseline ablation 4/4 MI WINS (Marjieh Study 1A harmonic: baseline anti-predicts consonance ρ = −0.71 vs MI stumpf_fusion +0.81). Frozen engine pin (bit-identical to paper-time freeze): • Commit: 318eb2f529d7103e8b7d80b01228357fdc4e0217 • Aggregate SHA-256: 482ade45c50f5d3bf5c90c122e495b2c3230e6e6edc6542f72f22e3b5da37f88 (Re-verifiable via the canonical find+shasum command documented in README.) Engine IP note: The MI engine source code remains under SRC9 Sonic IntelligenceLLC copyright and is NOT included in this deposit. The reproduction archive isan artefact-only release — all numerical results are reproducible from thedeposited cache substrate, datasets, and verification scripts without engineaccess. Research-use source access available upon email request toamace@bu.edu. Heavy artefacts vendored separately on Zenodo: • 138 GB MI engine pre-compute (per-frame .npz feature caches) • ~3 GB Gen-2 behavioural dataset metadata • 702 MB ds002725 mech×region BOLD intermediatesThe ~185 MB cache substrate (_unit_test_oracles) needed for reviewer-modeR³/T³ pytest is included here. Runtime: ~55 min on MacBook Air M2 8 GB cache-mode (~1 h 45 min live-enginemode). Bundle integrity: 1,097 verdicts re-verifiable via verify_full_ledger.py. Source code repository: https://github.com/amacerdem/Musical_Intelligence_Results Update (2026-05-31) — Note on bundled upstream datasets. This archive includes copies of several upstream research datasets under Musical_Intelligence_Results/datasets/ and one file under engine_outputs/consonance/ — namely PMEmo, ChillsDB, TenseMusic, DEAM, Emotify, Eerola film soundtracks, Marjieh 2024, Harrison 2024 (carillon), and Bowling 2018 — bundled for reviewer reproducibility convenience. Each bundled dataset remains the intellectual property of its original authors and is included under its own original license. See DATASETS_LICENSE.md in the archive root for attribution and authoritative-source URLs. The PolyForm Noncommercial 1.0.0 license of this archive applies only to MI-authored content and does not override or extend to any bundled upstream dataset. For commercial use of either MI content or any upstream dataset, contact the respective copyright holders. If you are an upstream author and prefer your dataset to be removed from a future version of this deposit, please email amace@bu.edu — removal will be prompt and acknowledged.
