# Scoring with an API judge (OpenAI, Gemini, a `vllm serve` endpoint). No GPU needed.
numpy>=2.0
scipy>=1.13
datasets>=3.0
json_repair>=0.30
