Anthropic introduced FlashProve, a research prototype that checks selected AI outputs against a model’s committed weights without revealing those weights. Its demonstration with Llama 4 Maverick requires roughly 70 to 95 times the GPU time of the reference FP8 inference on vLLM, while each proof is about 19 MB and can be verified on a CPU. The proof confirms computation performed by the committed model but does not establish the model’s authenticated identity or the factual accuracy of the answer. The current version is limited to 1,024 tokens, is not a feature for Claude customers, and remains an unmaintained research artifact without formal proof that the executable code matches the protocol or that zero knowledge holds for batched requests.
Source: Read the original article
