A new verification architecture for frontier AI training using zero-knowledge proofs could enhance governance frameworks by providing technical verification of training processes. This approach aims to make high-impact AI model assessments more reliable and enforceable, with a proof of concept expected in about 36 months.