Which prompt variant cites the source best?

Every night, our eval harness runs each prompt variant against the full hand-graded corpus. We score with ROUGE-L, BERTScore F1, and an LLM-judge sample on claim faithfulness. The top row is the current production default.