{
  "schema_version": "trec-rag-2026-rag-submission-bundle-v1",
  "bundle_id": "selected-evidence-sol-v1",
  "track": "TREC RAG 2026",
  "task": "RAG",
  "official_contract": {
    "checked_at": "2026-08-09",
    "guidelines_commit": "f281e88f61252662033c681df8b1ed2d0ceda97e",
    "test_data_commit": "a6255c10119a2984a874f46172d94045168ab1f3",
    "test_narratives_path": "trec-rag-data/trec-rag-2026/test-data/trec_rag_2026_queries.tsv",
    "test_narratives_sha256": "72dc2fd358d3eeda973397ccd7a8775545b19a6deaefc67709167eee6a9f8a2c",
    "expected_output_filename": "rag_output_trec_rag_2026.jsonl",
    "maximum_runs_per_task": 10,
    "evalbase_priority_required": true,
    "evalbase_organization_name": "NP Labs",
    "team_id": "NP",
    "submission_url": "https://ir.nist.gov/evalbase/conf/trec-2026"
  },
  "source": {
    "retrieval_artifact": "outputs/facet-deepseek-b40-v3",
    "generation_handoff_manifest_sha256": "31dc1b3578741339101b8199c0ea027e2d8405e828532aa58f86bf74c0542a0d",
    "topic_count": 119,
    "evidence_group_count": 691,
    "merged_multistage_runner_commit": "25201da2839220e941b986b3f1d04dba36890fed",
    "production_generation_commit": "736ececc5aed946df58c2fb113154a25f0b0bc0f"
  },
  "runs": [
    {
      "name": "multistage",
      "suggested_priority": 1,
      "run_id": "rag26-ms1-final",
      "run_desc": "Selected-evidence retrieval with bounded multi-stage Sol and Luna answer generation.",
      "path": "multistage/rag_output_trec_rag_2026.jsonl",
      "evalbase_response_path": "multistage/EVALBASE.md",
      "bytes": 1089595,
      "line_count": 119,
      "sha256": "ff384a62069062709abfe9cbfbcce2ba89c4f63ad857b917101488296e11677a",
      "provider_call_count": 1158,
      "provider_calls_by_model": {
        "openai/gpt-5.6-luna": 894,
        "openai/gpt-5.6-sol": 264
      },
      "provider_reported_cost_usd": "50.423054730",
      "failure_count": 0,
      "warning_count": 6,
      "warnings": {
        "planner_deadline_fallback_topics": [
          "rag2026-2",
          "rag2026-12",
          "rag2026-13",
          "rag2026-19",
          "rag2026-25",
          "rag2026-65"
        ],
        "operation_screen_crash_fallback_count": 0
      },
      "minimum_answer_words": 531,
      "maximum_answer_words": 1011,
      "minimum_reference_count": 14,
      "maximum_reference_count": 136
    },
    {
      "name": "singlepass",
      "suggested_priority": 2,
      "run_id": "rag26-ss1",
      "run_desc": "Selected-evidence retrieval with single-pass GPT-5.6 Sol answer generation.",
      "path": "singlepass/rag_output_trec_rag_2026.jsonl",
      "evalbase_response_path": "singlepass/EVALBASE.md",
      "bytes": 606888,
      "line_count": 119,
      "sha256": "91f757eb12291c118fe21e0828f0e607bba90460616e3405bd88f62eb67ab808",
      "provider_call_count": 121,
      "provider_calls_by_model": {
        "openai/gpt-5.6-sol": 121
      },
      "provider_reported_cost_usd": "52.14302500",
      "semantic_retry_topics": [
        "rag2026-52",
        "rag2026-76"
      ],
      "failure_count": 0,
      "warning_count": 0,
      "minimum_answer_words": 203,
      "maximum_answer_words": 749,
      "minimum_reference_count": 10,
      "maximum_reference_count": 40
    }
  ],
  "validation": {
    "autojudge_spec": "rag26",
    "autojudge_result": "PASS",
    "topic_count_per_run": 119,
    "exact_topic_order": true,
    "exact_narratives": true,
    "allowed_climbmix_citation_domains": true,
    "exact_hint_citation_rules": true,
    "unique_references_per_record": true,
    "maximum_answer_words": 1024,
    "organizer_schema_valid": true,
    "privacy_review": {
      "contains_required_official_narratives": true,
      "contains_generated_answer_text": true,
      "contains_reference_document_ids": true,
      "contains_selected_evidence_passage_text": false,
      "contains_raw_provider_responses": false,
      "contains_prompts": false,
      "contains_credentials_or_authorization_headers": false,
      "contains_qrels_gold_nuggets_or_evaluation_scores": false
    }
  }
}