Writing criteria the model reads the way you meant. Picking a threshold from what being wrong costs. Taking one judgement out of your LLM without breaking anything downstream.