<p>The rapid integration of generative AI into music creation has sparked a profound attribution crisis. Trained on vast corpora of human works, these models replicate stylistic signatures and sonic textures with remarkable fidelity, yet provide no traceable record of creative lineage. This opacity enables uncompensated appropriation, fuels economic displacement of human artists, and erodes moral rights in musical identity. This paper introduces the Compositional Provenance Framework (CPF), a sociotechnical architecture embedding forensic transparency into the generative pipeline. CPF includes three components: retrieval-augmented influence tracking, using a two-stage process–retrieving candidate sources via approximate nearest neighbor search, then computing gradient-based influence on this subset–to quantify training-data contributions at the token level; multi-dimensional fingerprinting with Optimal Pitch Alignment mapping generated segments to sources via chroma, timbral, and melodic metrics, robust to adversarial transformations; and an immutable provenance database with cryptographically-enforced governance logging attribution metadata while protecting proprietary data. Validated on a GTZAN dataset prototype, CPF achieves 100% retrieval accuracy under four adversarial conditions like tempo changes, pitch shifts, and noise injection. Despite advances, CPF has limits: reliance on training-data access, vulnerability to poisoning, influence underestimation in real-world noise, and limited portability to diffusion or hybrid architectures. Probabilistic scores risk misinterpretation and gaming. CPF marks the frontier of ethical generativity, making creative borrowing visible and contestable while recognizing sociotechnical bounds on perfect provenance in music. Code is publicly available to facilitate reproducibility.</p>