The Hallucination Defense
Summary
The article argues that logs alone cannot prove who authorized an action in multi-agent AI systems and proposes independently verifiable authorization artifacts called warrants. It introduces the concept of warrants (cryptographic, scoped, time-bound) and receipts to bind humans to actions and provide durable evidence across hops, improving accountability and reducing risk in automated business processes.