ModelProof: Bit-Exact, Permanently Stored AI Models You Can Verify in a Browser Tab

The project, built for the HackerNoon Decentralize AI Hackathon, ships a four‑step solution: MPW1 integer‑only inference (int8 weights, int32 biases, ReLU activations, arithmetic‑right‑shift requantization), a JSON manifest that records SHA‑256 hashes, size, Arweave transaction IDs, preprocessing rules, test vectors and an output digest, and finally immutable storage of both manifest and weights on Arweave via ANS‑104 data items tagged “App‑Name: ModelProof”. The demo uses a tiny 4‑8‑3 multilayer perceptron trained on the Iris dataset, achieving 100 % holdout accuracy before and after quantization; the weight file is only 117 bytes (TX SVHIFB9pR84SAv5TYwo_JUXJuAG8h7c1lTl‑wmVmHl0) and the manifest lives at TX XCnP6FwdPqVPz5REcI7hHmBR8X75bKdOMGWXTOstbLI. Users can paste the manifest ID into a web app hosted on Arweave, which downloads the weights, validates the hash, runs every test vector in both Python (standard‑library only) and JavaScript, and confirms the output digest matches. No server‑side component is involved, and the verification code is open‑source under MIT.

The effort addresses a persistent pain point in AI: model provenance is fragile because repositories can be altered or deleted, and floating‑point inference varies across CPUs, GPUs and libraries, making “same model, different result” disputes common. Existing verifiable‑inference proposals either tolerate numeric drift or rely on heavyweight zero‑knowledge proofs that are impractical for real‑world networks. By forcing inference into an exact integer domain, ModelProof eliminates drift and enables a simple hash‑based fingerprint of behavior, a technique already used for quantized models on mobile and edge devices. This positions the project as a pragmatic middle ground between ad‑hoc reproducibility checks and cryptographic proof systems, and it could become a reference implementation for anyone needing immutable, auditable model distribution—particularly in decentralized AI marketplaces or regulatory environments.

Scaling the approach will be the next hurdle. While the integer format works for the Iris MLP, larger models will require chunked weight files and Merkle‑root verification, a feature slated for the roadmap. Even if the storage and verification pipelines mature, ModelProof only guarantees that the uploaded model matches its manifest; it cannot attest to the honesty of the training data or claimed performance metrics. Consequently, malicious actors could still publish a deliberately flawed model and lock it in perpetuity. Watch for the upcoming Merkle‑root implementation, community‑driven “reproduced by” attestations on Arweave, and any integration with GPU‑backed verification services like Nosana, which will reveal whether the concept can survive the performance and security demands of production‑scale AI.