https://notes.karlmcguinness.com/tags/evals/Recent content in Evals on Control Plane by Karl McGuinnessHugoen-uspublic@karlmcguinness.com (Karl McGuinness)public@karlmcguinness.com (Karl McGuinness)Tue, 09 Jun 2026 09:00:00 -0700https://notes.karlmcguinness.com/notes/from-written-intent-to-executable-authorization/Tue, 09 Jun 2026 09:00:00 -0700public@karlmcguinness.com (Karl McGuinness)https://notes.karlmcguinness.com/notes/from-written-intent-to-executable-authorization/Microsoft’s ASSERT turns written behavior requirements into executable evaluations, making intent executable for verification. That answers what the agent did, not what it should have been allowed to do, and those are different controls: verification is detective, authorization is preventive, and for irreversible actions only the preventive one helps. Mission shaping is the preventive twin. It transforms an open-ended instruction into a bounded, machine-readable mission object that authorization derives from and enforces before the agent acts. A mission object is the authorization equivalent of an eval, with the same lineage from natural-language intent and the same dependence on a model to produce it, but a different lifecycle, a higher bar, and teeth an eval does not have. Together they close a loop where the same human goal drives both the boundary and the test, and the evals catch when the boundary was wrong.