When Large Language Models are used as judges for decision-making across various sensitive domains, they consistently exhibit unpredictable and hidden measurement biases, making their verdicts unreliable despite common prompt engineering practices.