[
  {
    "id":"maestro","title":"Enabling Factorized Piano Music Modeling and Generation with the MAESTRO Dataset","version":"arXiv:1810.12247v5 · 2019","url":"https://arxiv.org/html/1810.12247v5","kind":"Music paper",
    "claims":[
      {"id":"MA1","claim":"The system separates transcription, symbolic music generation and audio synthesis.","quotes":["three separately trained modules"],"where":"§1, paragraph after Eq. (1)","url":"https://arxiv.org/html/1810.12247v5#S1.p8.1","limit":"A particular piano architecture, not a requirement that every music system use these three models.","use":"Keep MIDI reasoning and sound rendering separate initially."},
      {"id":"MA2","claim":"MAESTRO's split criteria exclude a composition from appearing across splits.","quotes":["No composition should appear in more than one split."],"where":"§3.2 Dataset splitting","url":"https://arxiv.org/html/1810.12247v5#S3.SS2","limit":"This does not establish that Cadence's test pieces were absent from its inherited pretraining.","use":"Preserve composition/recording/duplicate-group splits and disclose pretraining uncertainty."},
      {"id":"MA3","claim":"The approximately 3 ms figure describes MIDI–audio dataset alignment, not online transcription latency.","quotes":["aligned with ≈3 ms accuracy"],"where":"§3; alignment procedure §3.1 and appendix","url":"https://arxiv.org/html/1810.12247v5#S3.p2.1","limit":"Calling the whole-recording alignment procedure offline is our characterization; the paper does not present this number as live inference delay.","use":"Qualify any future audio frontend separately."},
      {"id":"MA4","claim":"Its Music Transformer generates performance MIDI autoregressively.","quotes":["generates piano performance MIDI autoregressively"],"where":"Figure 1 caption; validation NLL in Table 7","url":"https://arxiv.org/html/1810.12247v5#S1.F1","limit":"Autoregressive modeling and validation likelihood are not demonstrations of a responsive live partner.","use":"Do not reject maximum-likelihood training simply because the product is interactive."}
    ]
  },
  {
    "id":"amt","title":"Anticipatory Music Transformer","version":"arXiv:2306.08620v2 · 2024 revision","url":"https://arxiv.org/html/2306.08620v2","kind":"Music paper",
    "claims":[
      {"id":"AM1","claim":"The method asynchronously conditions on supplied musical controls.","quotes":["conditioned asynchronously"],"where":"Abstract; §3 control-sequence definition","url":"https://arxiv.org/html/2306.08620v2#abstract1","limit":"Supplying a known control sequence is different from knowing an improvising player's future notes.","use":"Separate known instructions from unobserved future human performance."},
      {"id":"AM2","claim":"Control events can be serialized a chosen lead time before their associated musical events.","quotes":["δ seconds in advance"],"where":"§3 opening paragraph, before Definition 3.1","url":"https://arxiv.org/html/2306.08620v2#S3.p1.1","limit":"This serialization parameter is not a measured end-to-end MIDI or audio response time.","use":"Make availability and musical event time explicit rather than treating token order as wall-clock time."},
      {"id":"AM3","claim":"The discussion distinguishes causal filtering from bidirectional smoothing of controls.","quotes":["filtering (i.e., causal)","smoothing (i.e., bidirectional)"],"where":"§3 opening paragraph; two separated excerpts","url":"https://arxiv.org/html/2306.08620v2#S3.p1.1","limit":"Neither phrase establishes that the evaluated system follows an unknown live player continuously.","use":"Audit the actual information supplied at every decision."},
      {"id":"AM4","claim":"Anticipatory models are trained using maximum likelihood.","quotes":["standard maximum likelihood estimation"],"where":"§3.3 Training Anticipatory Models","url":"https://arxiv.org/html/2306.08620v2#S3.SS3.p1.1","limit":"The paper's infilling/continuation results are not a general guarantee of closed-loop musical quality.","use":"Start with a correctly specified supervised causal task; do not assume RL is mandatory."}
    ]
  },
  {
    "id":"streammuse","title":"Real-Time Language Model Jamming: A Case Study for Live Music Accompaniment Generation","version":"arXiv:2606.11886v1 · StreamMUSE · 2026","url":"https://arxiv.org/html/2606.11886v1","kind":"Music systems paper",
    "claims":[
      {"id":"SM1","claim":"StreamMUSE uses a discrete scheduling tick; §III-A describes quarter-beat ticks.","quotes":["tick to refer to the temporal/scheduling unit"],"where":"§III-B Data Representation; definition also §III-A","url":"https://arxiv.org/html/2606.11886v1#S3.SS2","limit":"A scheduling tick is not the same thing as measured wall-clock latency or waveform resolution.","use":"Keep one shared time reference, without automatically adopting this tick representation."},
      {"id":"SM2","claim":"The human and model streams are interleaved for autoregressive prediction.","quotes":["combined using an interleaving technique"],"where":"§III-B, before the InterL equation","url":"https://arxiv.org/html/2606.11886v1#S3.SS2","limit":"A multistream formulation is not itself proof that Cadence's forecast checkpoint accepts this different training contract.","use":"Distinguish human observations and actual generated history explicitly."},
      {"id":"SM3","claim":"The quoted latency ranking concerns deployment/network configurations, not a causal-versus-offline model comparison.","quotes":["highest base latency"],"where":"§V-B RTT Latency Characteristics","url":"https://arxiv.org/html/2606.11886v1#S5.SS2","limit":"The remote-server configuration is the subject of this phrase. Do not cite it as evidence that causal architectures inherently add latency.","use":"Measure network, inference and playback separately in our deployment."},
      {"id":"SM4","claim":"The paper discusses training/inference mismatch and a generation gap.","quotes":["known as exposure bias"],"where":"§III-C; revisited in §VI-C","url":"https://arxiv.org/html/2606.11886v1#S3.SS3","limit":"This is a practical challenge, not a theorem that cross-entropy cannot produce interactive music.","use":"Test generated-history feedback and real interruptions; do not qualify the model from teacher-forced scores alone."}
    ]
  },
  {
    "id":"streaming","title":"Streaming Generation for Music Accompaniment","version":"arXiv:2510.22105v1 · 2025","url":"https://arxiv.org/html/2510.22105v1","kind":"Music paper",
    "claims":[
      {"id":"SG1","claim":"Future-visibility and chunk-size combinations are studied through a grid of trained models.","quotes":["training a grid of"],"where":"§4 Experiment Setup, opening","url":"https://arxiv.org/html/2510.22105v1#S4","limit":"The results do not establish an arbitrary inference-time visibility knob for one unchanged checkpoint.","use":"Match the training information contract to the intended runtime."},
      {"id":"SG2","claim":"The tested low/future-context regime is judged insufficient for coherent accompaniment.","quotes":["insufficient for coherent accompaniment"],"where":"Abstract","url":"https://arxiv.org/abs/2510.22105v1","limit":"This is an empirical conclusion about the tested setup, not an impossibility result for all causal models or losses.","use":"Treat musical usefulness without future human input as an experiment, not an assumption."},
      {"id":"SG3","claim":"The paper's signed-visibility explanation conflicts with its printed feasibility inequality.","quotes":["predicts ahead to compensate for system latency"],"where":"§3.2 versus §3.4","url":"https://arxiv.org/html/2510.22105v1#S3.SS2","extra_url":"https://arxiv.org/html/2510.22105v1#S3.SS4","limit":"§3.2 assigns prediction-ahead to t_f<0; §3.4 prints t_f≥τ_sys+τ_jitter with positive delays. We do not copy this sign convention.","use":"Define our positive lead time directly from scheduling deadlines."},
      {"id":"SG4","claim":"The codec-delay equation and stated numeric value disagree in §4.2.","quotes":["of computation delay"],"where":"§4.2 Model Architecture","url":"https://arxiv.org/html/2510.22105v1#S4.SS2","limit":"The printed (N_q−1)/50 with N_q=4 equals 0.06 s, not the stated 0.08 s. This arithmetic check is ours.","use":"Do not import that number as a verified latency budget."}
    ]
  },
  {
    "id":"moshi","title":"Moshi: a speech-text foundation model for real-time dialogue","version":"arXiv:2410.00037v2 · 2024","url":"https://arxiv.org/html/2410.00037v2","kind":"Speech evidence / music analogy",
    "claims":[
      {"id":"MO1","claim":"Moshi factorizes modeling across time and across codec depth.","quotes":["Temporal Transformer","Depth Transformer"],"where":"§3.4.1, Eq. (1)–(3); separated excerpts","url":"https://arxiv.org/html/2410.00037v2#S3.SS4.SSS1","limit":"Speech codec factorization is not evidence that Cadence needs a new architecture.","use":"Keep as a later efficiency analogy, not a required rewrite."},
      {"id":"MO2","claim":"Inner Monologue supplies aligned textual content before audio-token generation.","quotes":["time-aligned text tokens"],"where":"§1 contributions; mechanism §3.4.4","url":"https://arxiv.org/html/2410.00037v2#S1","limit":"An aligned speech-text stream is not a demonstrated musical planner or harmony representation.","use":"Any symbolic musical planning stream would be a new hypothesis to test."},
      {"id":"MO3","claim":"The ASR/TTS experiment trains variants with opposite text/audio delay directions.","quotes":["train a streaming TTS model"],"where":"§5.7 Metrics; conceptual account §3.4.4","url":"https://arxiv.org/html/2410.00037v2#S5.SS7","limit":"Same framework, loss and data do not mean one unchanged set of weights provides three music products via a runtime switch.","use":"Make task-specific training and evaluation explicit."},
      {"id":"MO4","claim":"Mimi's 12.5 Hz figure is a latent codec frame rate.","quotes":["At 12.5Hz"],"where":"§3.3.1 Quantization rate","url":"https://arxiv.org/html/2410.00037v2#S3.SS3.SSS1","limit":"It is not an 80 ms quantization limit on every onset in the decoded waveform, nor Cadence's MIDI timing budget.","use":"Do not transfer speech-codec rates directly into music timing claims."},
      {"id":"MO5","claim":"At application inference, actual user audio replaces the model's predicted user stream.","quotes":["actual user audio is used instead"],"where":"§3.4.4 Inference of Moshi","url":"https://arxiv.org/html/2410.00037v2#S3.SS4.SSS4","limit":"This is a speech-system mechanism, not proof of musical responsiveness.","use":"Our analogous rule: predicted RH must never masquerade as newly observed human RH."}
    ]
  },
  {
    "id":"ghost","title":"The Ghost in the Keys: A Disklavier Demo for Human-AI Musical Co-Creativity","version":"arXiv:2511.01663v1 · Aria-Duet · 2025","url":"https://arxiv.org/html/2511.01663v1","kind":"Music demo paper",
    "claims":[
      {"id":"GH1","claim":"The demo uses an explicit handover and turn-based interaction design.","quotes":["the takeover signal","an embodied, turn-based interaction model"],"where":"§3.2 Real-time Engine; separated excerpts","url":"https://arxiv.org/html/2511.01663v1#S3.SS2","limit":"Its handover design is not a qualification of continuous simultaneous LH accompaniment.","use":"Turn-taking is a useful separately named milestone, not a substitute for the target product."},
      {"id":"GH2","claim":"Aria-Duet is presented as an alternating musical dialogue.","quotes":["enabling an alternating musical dialogue"],"where":"§3.2 Real-time Engine","url":"https://arxiv.org/html/2511.01663v1#S3.SS2","limit":"The name duet alone does not establish continuously following a performer who keeps playing.","use":"Keep continuation, call-and-response and simultaneous accompaniment distinct."}
    ]
  },
  {
    "id":"aria","title":"loubb/aria-medium-base — model card","version":"Hugging Face main · inspected 18 Sep 2026","url":"https://huggingface.co/loubb/aria-medium-base","kind":"Model documentation",
    "claims":[
      {"id":"AR1","claim":"The model card recommends continuation from an existing MIDI prompt.","quotes":["generating continuations of existing MIDI files rather than generating music from scratch"],"where":"README, Usage Guidelines / Intended Use, line 26 in inspected copy","url":"https://huggingface.co/loubb/aria-medium-base/raw/main/README.md","limit":"Calling it our musical prior is our description; the card does not establish live-following capability.","use":"Retain the accepted pretrained musical foundation while changing the interaction task."},
      {"id":"AR2","claim":"The card discloses compositional memorization risk.","quotes":["have been compositionally memorized by the model"],"where":"README, Data Memorization Considerations, line 29","url":"https://huggingface.co/loubb/aria-medium-base/raw/main/README.md","limit":"No quantified rate is supplied. A new fine-tuning split does not erase inherited exposure.","use":"Avoid calling our held-out set pretraining-unseen without an overlap audit."},
      {"id":"AR3","claim":"The model repository declares Apache-2.0.","quotes":["apache-2.0"],"where":"README YAML frontmatter","url":"https://huggingface.co/loubb/aria-medium-base","limit":"This declaration is not commercial clearance for the training corpus or underlying compositions.","use":"Track model and data terms separately."}
    ]
  },
  {
    "id":"aria-midi","title":"The Aria-MIDI Dataset — dataset card","version":"V1 · Hugging Face main inspected 18 Sep 2026","url":"https://huggingface.co/datasets/loubb/aria-midi","kind":"Dataset documentation",
    "claims":[
      {"id":"AD1","claim":"The compiled MIDI dataset declares CC BY-NC-SA 4.0.","quotes":["cc-by-nc-sa-4.0"],"where":"README YAML frontmatter and License section","url":"https://huggingface.co/datasets/loubb/aria-midi/raw/main/README.md","limit":"This is the publisher's dataset declaration, not an assessment of every underlying musical work.","use":"Research/demo status is not automatic commercial clearance."},
      {"id":"AD2","claim":"The corpus originates from automatic transcription of public solo-piano recordings.","quotes":["transcribing publicly available solo-piano audio recordings into MIDI files"],"where":"README Dataset section, line 45 in inspected copy","url":"https://huggingface.co/datasets/loubb/aria-midi/raw/main/README.md","limit":"Transcribed notes are not error-free performance or hand-assignment ground truth.","use":"Retain data/provenance audits and inspect failure cases."},
      {"id":"AD3","claim":"The card identifies Pruned as the pretraining subset; Deduped has 371,053 files.","quotes":["used to pre-train Aria"],"where":"README subset callout and Download (V1) table","url":"https://huggingface.co/datasets/loubb/aria-midi/raw/main/README.md","limit":"Pruned contains 820,944 files. The card does not quantify overlap of our Deduped selection with a particular checkpoint's training.","use":"Do not confuse dataset subset names with proof of evaluation independence."}
    ]
  },
  {
    "id":"hands","title":"Piano-Hands — pinned source repository","version":"Commit c23c1b1f15bea2e5017b526af8166f318608877c","url":"https://github.com/asigalov61/Piano-Hands/tree/c23c1b1f15bea2e5017b526af8166f318608877c","kind":"Model/code documentation",
    "claims":[
      {"id":"PH1","claim":"The project is intended to split a mixed piano composition into hand parts.","quotes":["split mixed Piano compositions into separate parts for each hand"],"where":"Pinned README, line 2","url":"https://raw.githubusercontent.com/asigalov61/Piano-Hands/c23c1b1f15bea2e5017b526af8166f318608877c/README.md","limit":"The inspected short README supplies no validated accuracy benchmark. Our outputs remain predicted labels.","use":"Note preservation and implementation parity must not be presented as hand-label accuracy."},
      {"id":"PH2","claim":"The pinned code repository contains an Apache-2.0 license.","quotes":["Apache License Version 2.0"],"where":"Pinned root LICENSE","url":"https://raw.githubusercontent.com/asigalov61/Piano-Hands/c23c1b1f15bea2e5017b526af8166f318608877c/LICENSE","limit":"The code license does not establish training-data rights or reliability of predicted hand assignments.","use":"Separate software permission, data rights and scientific validation."}
    ]
  },
  {
    "id":"mido","title":"Mido — MIDI Objects for Python","version":"Stable documentation inspected 18 Sep 2026","url":"https://mido.readthedocs.io/en/stable/","kind":"Software documentation",
    "claims":[
      {"id":"MI1","claim":"Mido supplies MIDI message, port and file I/O.","quotes":["Python library for working with MIDI 1.0 ports, messages and files"],"where":"Documentation Overview, opening sentence","url":"https://mido.readthedocs.io/en/stable/","limit":"MIDI I/O is not the neural musical model, sound synthesizer or proof of a live accompaniment policy.","use":"Reuse it behind the score/MIDI boundary rather than confusing it with the learned prior."},
      {"id":"MI2","claim":"Mido's code license is MIT.","quotes":["The MIT License"],"where":"Repository LICENSE","url":"https://github.com/mido/mido/blob/main/LICENSE","limit":"This software license says nothing about music-data licensing.","use":"Keep implementation dependencies separate from corpus rights."},
      {"id":"MI3","claim":"Mido's documentation has a separate CC BY 4.0 declaration.","quotes":["licensed under the Creative Commons Attribution 4.0 International License"],"where":"Documentation License section","url":"https://mido.readthedocs.io/en/stable/","limit":"The documentation declaration is distinct from the MIT-licensed code.","use":"Cite the appropriate artifact, not a blended license claim."}
    ]
  }
]