Governance that stops at policy documents cannot reconstruct what a machine actually did.
A policy that exists only in a wiki is not evidence of execution. A chat log is not evidence of execution. A checkbox is not evidence of execution.
Autonomous systems fail in public when nobody can answer three questions after the fact. What was requested. Whether it was permitted. What actually occurred. If those answers live in someone’s memory or in a transcript of a model talking about itself, the organization does not have a record. It has a story.
Evidence after execution is not a luxury for regulated industries. It is how operators learn, how incidents are bounded, and how trust is rebuilt. The point is not to prove that an AI will always be safe. No serious system claims that. The point is a bounded question: was this action authorized, and what happened.
Hood publishes that distinction because the market is already asking for it. The world is giving software the ability to act. Infrastructure for trusted action has to include a record that can be replayed, not a hope that the model behaved.
Hood Research. Category commentary. Not a specification and not an implementation guide.