Scalable structures for verifiable audits beyond enclaves

Develop scalable approaches for verifiable audits of AI systems that do not require all evaluation and inference to occur inside trusted execution environments, thereby reducing computational overhead while maintaining strong verification guarantees.

Background

Zero-knowledge proofs and TEEs can enable trustless audits, but current designs may impose significant overhead and require enclave-only evaluation and inference, limiting practicality and scope.

Finding scalable audit verification structures would make trust-preserving audits feasible for a broader range of systems and use cases.