{ "dataset": "multihop", "questions": 300, "candidate_count_total": 6000, "candidate_pool": "Top-20 neural MiniLM retrieval from full 609-article corpus (no gold insertion)", "gold_support_count_distribution": { "3": 96, "2": 158, "4": 46 }, "gold_available_fraction": 51.67, "gold_available_count": 155, "graph_link_fraction": 4.0, "oracle_complete_coverage_pct": { "2": 35.0, "3": 48.0, "5": 51.67, "8": 51.67 }, "complete_coverage_pct": { "dense_cosine": { "2": 6.0, "3": 11.67, "5": 21.67, "8": 31.67 }, "crossencoder": { "2": 7.0, "3": 13.67, "5": 23.0, "8": 36.67 }, "dense_graph": { "2": 6.0, "3": 11.67, "5": 22.33, "8": 32.33 }, "crossencoder_graph": { "2": 7.67, "3": 13.67, "5": 26.0, "8": 36.67 } }, "crossencoder_graph_minus_crossencoder_at3_pp": 0.0, "ci95_pp": [ -1.33, 1.33 ], "paired_wins_losses_ties_at3": { "wins": 2, "losses": 2, "ties": 296 }, "timing_seconds": { "index_seconds": 8.61646041700078, "query_embedding_seconds": 2.9061220870007674, "total_prep_seconds": 11.600812000000587, "index_size": 609, "candidate_pool_size": 20, "crossencoder_scoring_seconds": 190.511, "crossencoder_scoring_ms_per_question": 635.04, "crossencoder_forward_seconds": 176.907, "graph_selection_total_s": 2.271, "graph_selection_ms_per_question": 7.57 }, "models": { "dense": "user uploaded all-MiniLM-L6-v2", "crossencoder": "user uploaded cross-encoder/ms-marco-MiniLM-L6-v2" }, "protocol": "MiniLM embeddings at max 128 WordPiece for dense. Cross-encoder truncates paired query+passage at 256 WordPiece. Both score exact same candidate pools. Fixed evidence graph .4 minmax relevance + .5 title in already-selected passage + .2 query-title score. No TF-IDF anywhere in candidate construction or scoring. Gold used solely for pool audit and evaluation." }