How can an AI agent decide which rules to follow? One rule permits an action, another may impose a condition, an exception, or a conflicting duty. Deterministic systems resolve explicit relationships, but implicit language rules can be missed. Refusing every unresolved action is safe yet blocks legitimate behavior. Our initial hypothesis was that geometric measurements could provide a basis for judgment. We encoded actions and policies as vectors and examined whether their geometry could reveal governing policies and interpret the action’s relation to them. Four studies showed these approaches failed to deliver reliable pre‑action judgment. In a final synthetic study, a lexical router recovered every governing and blocking policy and cut median policy checks by 97.7%. Nevertheless the pipeline escalated all 2,304 test actions, including those that should have been allowed. Supplying all policies to the same downstream component did not change the decision. Finding the policies did not solve the interpretation problem. We therefore revised the hypothesis: AI judgment requires a consequence graph that links the actor, authority, policy conditions, exceptions, and the changes an action would cause. Follow‑up work will test whether making these relationships explicit helps the agent distinguish when to act, stop, or request review.
Review