Independent assurance for AI agents. Audit-ready evidence for consequential agent actions. The system that performs the action should not vouch for the action.