Anthropic introduces FlashProve to cryptographically verify which model produced an answer
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Cryptographic confirmation of inference could make model selection and payment verifiable without requiring a lab to disclose its weights.
The provable inference system is designed to confirm that a particular answer came from the stated model. The lab publishes a digital model "fingerprint" that is matched to the output and cryptographically verified without revealing the weights. Users could detect a hidden switch to a cheaper model, while a third-party developer could prove that a task was actually performed by the specified system. Verification is designed to run on a local device. Anthropic published a technical specification and open-source code for selected components, emphasizing that this is an intermediate research result rather than a finished product.
Sources 1