A developer builds a RAG system and creates an eval suite with 100 test questions. They use Claude Opus 4.8 as an LLM judge to score answers on a 1-10 quality scale. After running the eval, they observe that 95% of scores are between 6 and 8, with very few scores at the extremes (1-3 or 9-10), even though some answers are clearly excellent and others are clearly poor. What is the most likely cause and how should the eval be improved?