Anthropic has launched and open-sourced FlashProve, a provable reasoning system for large models. Built on zkSNARK technology, it verifies model outputs through a three-step process: offline generation of a unique commitment ID for the weights, online generation of cryptographic proofs of inference, and completion of verification in 17 seconds on an ordinary CPU. This process requires neither access to the weights nor rerunning the model, yet it can accurately verify whether a large model's output was generated by a specified target model—effectively preventing wrapper behaviors such as model swapping and output tampering, with security robust against quantum attacks. The system can support a 400-billion-parameter-level MoE large model with just a single H100, and it solves the problem of large VRAM usage through streaming processing. More than 95% of the project's content was generated by Claude, with full mathematical formal proofs completed via Lean4; humans only need to finalize the top-level security definitions, thereby ensuring the three core properties of completeness, knowledge soundness, and zero knowledge. The system also includes companion specifications designed specifically for agent operation, pioneering a new engineering paradigm in which AI builds systems and Lean guarantees authenticity.