Anthropic's FlashProve proves which model answered without revealing its weights: about 110 tokens/s on one H100, 19 MB proofs checked on a CPU.