Guide
An AI receipt answers one question: what exactly just happened?
An AI receipt is a tamper-evident record of a single action taken by an AI system. Not a log line, a receipt: it states what the system intended before acting, what actually resulted, cryptographic hashes of the inputs and outputs, which policy allowed or denied the action and why, and a hash link to the previous receipt so the whole chain breaks visibly if anyone edits history.
Anatomy of a receipt
Our production receipts carry eight fields: the intent before the action, the result after it, an input hash, an output hash, the policy decision with reasons, a flag when the action needs human review, a hash link to the previous receipt, and a correction reference when an entry fixes an earlier one. Two of those deserve emphasis. Denials get receipts too, because a record that only remembers successes is advertising. And corrections are new entries that point at what they fix, never silent edits, so honesty about mistakes becomes visible evidence of integrity instead of a liability.
How to read one in sixty seconds
Open the chain and pick any entry. Check the timestamp and the actor. Read the intent, then the result, and ask whether they match. Confirm the hash link points at the entry before it. That is the whole skill, and it is the same skill an auditor uses, which is why receipts anchor our federal governance approach and every tier of human-in-the-loop enforcement we run.
Receipts everywhere or nowhere
The discipline only works when it is universal: local AI hands, deploys, denials, even our public scoreboard follows the spirit, counters that never decrement. Want it in your systems? Start with checkable AI, then talk to the team.