Zero-Knowledge Proofs Could Verify Frontier AI Training Compute Claims

A proposed verification architecture combines pre-committed training specs, inter-node network observations, and on-the-fly Merkle commitments of intermediate computation, verified via zero-knowledge proofs — arguing current ZKP impracticality is an engineering constraint, not a fundamental one. This directly targets the self-reporting gap in compute-based governance frameworks like the EU AI Act and proposed international agreements.