How can I maintain evidence of what an AI agent actually executed?

Keep a record at the moment of authorization, not a chat log afterward. Hood records who acted, what was requested, whether it was permitted, and what happened, in a form that can be replayed. Transcripts and UI checkboxes are not evidence of execution authority.

Deployment scope is determined during evaluation. Hood does not prove that AI will always be safe.

If this problem matches your environment, START with the systems, effects and authority boundary you need to control.