{
  "project_test_rows": 10,
  "systems": {
    "final_constrained_rag": {
      "reader": "google/flan-t5-large",
      "reader_revision": "0613663d0d48ea86ba8cb3d7a44f0f65dc596a2a",
      "retriever": "sentence-transformers/all-MiniLM-L6-v2",
      "retriever_revision": "1110a243fdf4706b3f48f1d95db1a4f5529b4d41",
      "omop_target_accuracy": 1.0,
      "ragbench_hit_at_1": 0.7375510204081632,
      "ragtruth_hallucination_f1": 0.632,
      "mirage_hit_at_1": 0.6779100529100529
    },
    "base_closed_book_flan_t5_large": {
      "omop_target_accuracy": 0.5,
      "ragbench_hit_at_1": 0.0,
      "ragtruth_hallucination_f1": 0.518,
      "mirage_hit_at_1": 0.0
    },
    "rag_flan_t5_small": {
      "omop_target_accuracy": 0.3,
      "ragbench_hit_at_1": 0.7375510204081632,
      "ragtruth_hallucination_f1": 0.632,
      "mirage_hit_at_1": 0.6779100529100529
    },
    "rag_flan_t5_base": {
      "omop_target_accuracy": 0.6,
      "ragbench_hit_at_1": 0.7375510204081632,
      "ragtruth_hallucination_f1": 0.632,
      "mirage_hit_at_1": 0.6779100529100529
    }
  },
  "secondary_retrieval_metrics": {
    "ragbench_recall_at_3": 0.7549523809523809,
    "mirage_recall_at_3": 0.9129872134038801
  },
  "notes": [
    "All RAG reader variants share the same MiniLM retriever and groundedness policy, so their public component-benchmark values are identical.",
    "The final pipeline locks the structured target to the highest-ranked corpus metadata; FLAN-T5 writes the rationale only."
  ]
}