Reconciling intent has a bootstrap problem: it's inferred from the same model you're constraining, so it rationalizes. Side-effect gates — spend, irreversible writes — can't be talked around.