When One Answer Is Not Observable: Ambiguity-Preserving Inference for Polyphonic Music Analysis

This preprint presents an ambiguity-preserving framework for polyphonic music analysis and structural inference developed within WAVEMULATOR. The work studies a recurring failure mode in music information retrieval: observed acoustic or structural evidence may support multiple latent explanations even when a conventional system is forced to output a single source, identity, ownership relation, or trajectory. The study separates observed evidence from source-level commitment and evaluates harmonic/fundamental ambiguity, source multiplicity, weak-element/noise ambiguity, temporal identity, provenance accounting, and large-scale structural ambiguity. Controlled experiments preserve adverse results and label corrections rather than removing them after inspection. The research record includes an independent audit against 1,399,331 saved graph records. In held-out harmonic conditions, false fundamental hypotheses decrease from 2,287 to 660 while misses decrease from 84 to 42. In fresh weak/noise conditions, supported fundamental-window evidence increases from 84.62% to 96.25% while false hypotheses decrease from 5,857 to 2,709. Temporal experiments expose a false-stitch/continuity frontier rather than claiming a solved tracker. A subsequent architecture-scale validation uses 63,355 unique real-audio-derived structural signatures. Across ten deterministic O₂–O₅ coverage rounds, 2,534,200 cases and 8,869,700 signature participations complete with zero post-stabilization invariant failures. All 63,355 signatures additionally participate simultaneously as competing hypotheses in each of 64 shared-evidence stress rounds, totaling 4,054,720 hypothesis participations. A four-factor cardinality stress stores 253,420 explicit alternatives while representing a theoretical 63,355⁴ configuration space without materializing Cartesian world objects. These experiments validate structural ambiguity handling, determinism, evidence accounting, and scaling behavior. They are not presented as exhaustive combinatorial enumeration, ground-truth copyright or plagiarism detection, authorship inference, or population-level music-recognition accuracy. DOI: 10.5281/zenodo.23220126