Note the benchmark: 8 subtasks, train/validation/test splits

Note decision-level substrate: 15 exhaustively annotated court decisions

Note predictions, LLM-judge decisions, and the exact prompts