Four things get confused in the 'AI accountability' conversation. They answer different questions. Only one leaves a portable, independently verifiable record of the action itself.
Built to debug software. Scattered across services, editable after the fact, and usually silent on whether the action was allowed.
Decide in the moment: block the risky output, route the uncertain one to a human. They act; they do not leave a portable, verifiable record of what they decided.
Periodic, document-driven, human-attested. Answers about the program, not about the specific action an agent just took.
Runtime-neutral. Every consequential action passes a deny-by-default gate; the verdict is sealed into a hash-linked Decision Receipt a third party can replay. Complements the guardrail: they decide in the moment; we prove it forever.