Published onOctober 7, 2026ReferenceApproval Boundaries for Expensive, External, and Irreversible AI ActionsHuman approval is useful only when it binds a specific actor, target, action, limit, and state. Classify tool risk and prevent vague consent from becoming broad authority.