{
  "runner_version": "kingy-document-runner/1.5.5",
  "questions_sha256": "5894838cffa707ecd654b2592b3a85cff9d6fcb6a262d99c9e68833e84f14f8d",
  "references_sha256": "a1779ef1ed7d518710545a8413d59822363d7403bf95331f4d9189bd8af78c2b",
  "runner_sha256": "ac8d1ccb48ebe017f7e2794392598967695ef3186fcf4d3889dfd579e8479458",
  "source_changed_during_run": false,
  "system_prompt_sha256": "80fa556d1117317a79713e3b25b150bf3e351e0a0587e8ce258166b2836b042e",
  "parse_seconds": 0.006117208999057766,
  "index_seconds": 0.00022616700152866542,
  "split": "selftest",
  "repetitions": [
    {
      "repetition": 1,
      "attempts": 4,
      "answerable_count": 3,
      "answerable_passes": 2,
      "unanswerable_count": 1,
      "correct_abstentions": 1,
      "citation_count": 4,
      "valid_citation_count": 3,
      "answer_correctness": 0.6666666666666666,
      "citation_precision": 0.75,
      "abstention_accuracy": 1.0,
      "qualified": false,
      "by_category": {
        "cross_document": {
          "attempts": 1,
          "passed": 0
        },
        "direct": {
          "attempts": 2,
          "passed": 2
        },
        "unanswerable": {
          "attempts": 1,
          "passed": 1
        }
      },
      "model_calls": 4,
      "model_request_count": 10,
      "guard_only_answers": 0
    }
  ],
  "complete_planned_evaluation": false,
  "qualified": false
}
