Anchoring Bias in LLM-as-a-Judge Systems: Prior Scores Compromise Evaluation