Nov 2, 2026
FlashProve is an open-source prototype for checking that an AI response came from a committed set of model weights without exposing the weights.
Anthropic and collaborators released code, a technical report and Lean proofs under Apache 2.0. The prototype demonstrated proof generation for a quantized Llama 4 Maverick model, with substantial compute overhead. Its proofs concern the model used to generate an output, rather than the accuracy of that output. Research credits: Roy Rinberg, Alex Allain, Alessandro Chiesa, Miranda Christ, Rahul Dalal, Giacomo Fenzi, Ben Mayer and Guy Weissenberg.