Anthropic's FlashProve verifies which AI model wrote a response

Anthropic's open-source FlashProve generates zkSNARK proofs at 110 tokens per second for Llama 4 Maverick, verifying which model wrote a response — but only for short, deterministic outputs.