Anthropic researchers have published a design for FlashProve, an open source scheme they describe as provable inference, meaning a cryptographic guarantee that a particular response was produced by a particular model.
The metaphor they reach for is a signed email. A signature lets the recipient trust that a message really came from whoever claims to have sent it. FlashProve applies that trust pattern to model output, so a user can check whether the assistant they are talking to is the one advertised.
The adversary here can be the vendor. A lab could quietly serve a cheaper model to protect margins, or a costlier one to make its benchmark numbers look stronger. Outsiders make similar claims: a developer insisting that a frontier security model reviewed a codebase, or an executive saying a strategy document got the nod from some advanced system.
Getting there requires labs to publish a fingerprint for each model they ship. Those fingerprints can later be matched against actual outputs and verified. They do not expose weights, the internal numbers that drive behavior and which closed vendors guard jealously, so publishing costs a lab nothing it regards as valuable.
Verification is also built to stay modest. The design contemplates checking a proof on an ordinary local device rather than a data center.
Anthropic put out a technical specification along with working code on GitHub under an open source licence. The authors describe the post as intermediate progress rather than a finished research agenda, and they say nothing is planned for deployment at Anthropic or anywhere else. The stated hope is that outsiders pick the work up and improve it.
