LLM-as-a-judge is reliable, but only when you control for its known failure modes. Used as a raw scorer it inherits three documented biases…