How it works

1. An AI produces an output: a summary, a decision, a piece of code.
2. By design, a receipt is sealed with it, recording the exact evidence the output rested on, the rule applied, and who issued it.
3. The standard lets an independent party you authorise check it. On the published test cases, a public reference checker reruns each check offline, from the record and the evidence it cites, and reaches the expected answer without contacting us or the producer.
4. Designed so a change reaches what was built on it. In the standard, when the state of a cited record changes, every record that cites it shows the change the next time it is checked, to the depth the checker follows.

We say tamper-evident, not tamper-proof: any change to a record is designed to be detectable and provable by an independent party you authorise.

You can run a check yourself on the published test cases: see Try a check. The rules a check follows are published as a standard: see The standard.