Checkable Claims About AI Inference and Compute Use Without Trusted Hardware

People who use AI, and governments that want to pace it, are asked to trust claims about which model answered and how a data center's chips were used.

The paper describes a way to check those claims with zero-knowledge proofs, a sealed log and random checks on ordinary rented machines, with no trusted hardware, plus light-speed timing that is designed but not built.

It separates what has run from what is designed: a zero-knowledge proof of a 7B model runs end to end, but a verifier flaw found on 29 Sep is fixed and earlier proofs are being re-checked, and a faster GPU prover does not hide anything yet.