mattmireles/mrt2-three-clocks-evidence

MRT2 Three Clocks — System Paper Evidence

Raw device evidence backing Throughput Is Not Liveness: Three Clocks for GPU-Free Music Generation on iPhone (Mireles, 2026).

Every file here is hash-bound to a machine-readable receipt in the public paper repository: https://github.com/mattmireles/magenta-realtime-2-iphone under validation/results/system-paper/. Re-hash any file in this dataset with sha256sum and compare against the corresponding manifest to verify it has not changed since publication.

Layout

• context12/ — the corrected A17 Pro sustained-throughput run (610 s foreground capture): PCM WAV, event trace, token capture, and summaries. Backs a17pro/context12/context12-soak-manifest.json.
• crossover/ — three seeds (20260718, 271828, 1618033) of the 2x2 token-source x decoder-path crossover, plus the FLOAT32 pre-iSTFT split, corrected-DSP controls, and 12-frame context arms. Backs crossover/seed-*.json and crossover/aggregate.json. decoder-context-probe-tokens.npy is the token file underlying crossover/decoder-context-probe.json.
• liveness/reset-factorial/ — all 12 arms (3 seeds x {off, kv-only, feedback-only, both}) of the frozen unrefreshed-liveness gate: WAV plus codebook-local token capture per arm. Backs liveness/g5-manifest.json.
• liveness/judge-events/ — the 9 event-centered model-vote lineups (sealed clip mapping + vote result + the three candidate/baseline/corruption-control clips) that back the listening checks in g5-manifest.json. These are automated model votes, not human-subject data.
• steering/ — the final post-ring steering run: WAV, event trace, token capture, paired-latency detector report, and device launch logs. Backs steering/g6-manifest.json.

What is intentionally not included

• The signed, compiled iOS application binary (its hash is recorded in the manifests as signedExecutableSha256 / executable, but the binary itself is the shipping product and stays private).
• Exploratory pilot/smoke runs and abandoned configurations that never became a numbered result in the paper.
• The single blinded human engineering listening check: it is not machine-hash-bound in any manifest and is reported only as prose in the paper's limitations.

License

The underlying mrt2_small checkpoint is CC-BY-4.0 (Google DeepMind). Audio and token captures derived from it are released under the same license.