Overview
- Anthropic published FlashProve as open‑source code and formal artifacts so providers can commit to model weights and supply a proof that a given response came from that committed model.
- The team demonstrated feasibility by proving outputs from Meta’s 400B‑parameter Llama 4 Maverick on a single NVIDIA H100, reporting roughly 110 tokens per second for proof generation and about 17 seconds to verify a 1,024‑token proof on a 24‑thread CPU.
- Each 1,024‑token proof is about 19 MB and proof generation currently uses roughly 70–95 times more GPU time than normal inference, making the system costly and limited to integer‑quantized models, greedy decoding, and a 1,024‑token window.
- FlashProve proves provenance only, not answer correctness, and Anthropic released C++ and Lean 4 verifiers plus machine‑checkable proofs under Apache 2.0 while noting the work remains a prototype that needs engineering, audits, and broader decoding support.
- The project targets real supply‑chain trust problems by giving auditors and clients a technical way to verify which model produced an output, but practical adoption will depend on cutting proof cost, expanding features, and independent review of verifier implementations.