Verifiable Deep Learning: Smart Contract--Enabled Model Integrity and Provenance on Distributed Ledgers

Deep learning artifacts undergo long and mutable lifecycles in which datasets, preprocessing pipelines, checkpoints, evaluation configurations, and deployed versions may be controlled by different actors. Hash anchoring makes individual bytes tamper-evident but does not establish that a checkpoint was derived from authorized data or that its alleged computation actually occurred. We present ProofWeave, a provenance-coupled verification specification that treats model evolution as a typed, policy-constrained transition system. Each transition binds content commitments, authenticated parent references, version rules, and explicit evidence obligations; the proposed contract state machine admits only policy-valid transitions. We distinguish full computational proofs, conditioned on a sound external proving backend, from probabilistic committed-step audits with a finite-population detection bound. For transparent evaluation we implement a local integrity prototype, not a live blockchain or zero-knowledge proof service. Seven deterministic integrity fault types are assessed across three public datasets; four nested validators cover 2, 4, 6, and 7 of the nine broader threat classes discussed. To expose an essential security boundary, we additionally perform 105 real MLP training runs over seven random seeds and five controlled training-label corruption rates. Hash/signature microbenchmarks and Merkle trees of up to 65,536 leaves characterize local scaling. A controlled Monte Carlo audit study quantifies probabilistic detection under committed-step sampling, separately from the local measurements. The evidence supports preliminary provenance-check correctness and quantifies its limits; it does not establish on-chain costs or end-to-end verifiable neural training.