Ask it to do something it should not do.

warrant sits between an agent and the systems it would otherwise reach. Type an action the way an agent would propose one. Nothing here is connected to a real system, and nothing is kept between requests.

propose an action
the record notices when you edit it

Three decisions are sealed into a chain, then one of them is rewritten directly in the database and its own hash resealed to match, which is what someone covering a track would do. Verification runs before and after.

Thirteen rules, each citing a real authority. Raise only, so a Class 2 match beats a Class 1 match on the same action. Fail closed, so an action matching no rule is Class 2 rather than assumed safe. Source and the two-arm evaluation: github.com/mariaangelikabuilds/warrant