FlashProve tests whether an AI reply came from committed model weights
Anthropic’s open research prototype connects zero-knowledge proofs with model provenance, while leaving substantial implementation and decoding limits.
Anthropic released FlashProve on November 2 as an open-source research prototype for provable inference. A provider commits to model weights and supplies a cryptographic proof linking a response to those weights, without revealing them. The release includes protocol documentation, executable code and security statements formalized in Lean.
Its reported evaluation uses a quantized Llama 4 Maverick model and contexts of up to 1,024 tokens. The proof concerns an integer implementation with greedy decoding; it does not automatically cover floating-point execution or sampled responses. Anthropic also separates its Lean guarantees from its faster executable verifier, whose equivalence is not formally guaranteed.
For a team buying model access through intermediaries, provenance could become a useful acceptance check: did the contracted computation produce this reply? It does not establish that the reply is factually correct, unbiased or suitable for a decision. The practical next question is whether the exact model, decoding behavior, context length and verification cost fit the intended workflow.