{
  "schema": "invisiblebench-score-evidence/v1",
  "benchmark_version": "4.0.0",
  "result_contract_version": "2.1.0",
  "generated_at": "2026-07-10T16:41:05.831369+00:00",
  "profile": "publish",
  "judge_model": "openai/gpt-5-mini",
  "model_count": 4,
  "scenario_count": 63,
  "check_count": 50,
  "row_count": 252,
  "mode_result_count": 12600,
  "claim_posture": "research_evidence_only",
  "claim_ready_check_count": 0,
  "content_notice": "Automated check verdicts over synthetic benchmark conversations. Model outputs may be unsafe or hallucinated. No check is a comparative Safety claim unless claim_ready; Care results are directional. This is research evidence, not medical, legal, crisis, or caregiving advice.",
  "source_scan_sha256": "b93f1331f8cd2c8c4e8692fedb3b6c675b4e9b49d992791e91486561e25496dc",
  "source_merge": {
    "schema": "invisiblebench-scan-merge/v1",
    "benchmark_version": "4.0.0",
    "result_contract_version": "2.1.0",
    "generated_at": "2026-07-10T16:40:49.186544+00:00",
    "profile": "publish",
    "judge_model": "openai/gpt-5-mini",
    "model_count": 4,
    "scenario_count": 63,
    "row_count": 252,
    "actual_cost_usd": 24.19056412999998,
    "actual_billable_api_calls": 9190,
    "output_file": "per_run.jsonl",
    "output_sha256": "b93f1331f8cd2c8c4e8692fedb3b6c675b4e9b49d992791e91486561e25496dc",
    "sources": [
      {
        "artifact_id": "20260710_163537",
        "file": "per_run.jsonl",
        "sha256": "0dcd9e1a8c5c02fc89036a096083250f7fc5995ccc8f7dd3c191dbd0712ef1fa",
        "row_count": 63,
        "model_ids": [
          "anthropic/claude-opus-4.8"
        ],
        "transcript_source_artifacts": [
          "run_20260709_135838"
        ],
        "actual_cost_usd": 5.8523027999999995,
        "actual_billable_api_calls": 2300
      },
      {
        "artifact_id": "resolved_20260710_164006",
        "file": "per_run.jsonl",
        "sha256": "3fa5bac1424a8e31910fc3fee924982c5a47d8e158352d4ff7710b008c3633e6",
        "row_count": 63,
        "model_ids": [
          "google/gemma-4-31b-it"
        ],
        "transcript_source_artifacts": [
          "run_20260709_010622",
          "run_20260709_012230"
        ],
        "actual_cost_usd": 5.972041909999997,
        "actual_billable_api_calls": 2309
      },
      {
        "artifact_id": "20260710_163042",
        "file": "per_run.jsonl",
        "sha256": "0eaa5dc80a8a6c01591e8ba876cbcd11bc0b91b9a5ae4ac545c1445aefdd674a",
        "row_count": 63,
        "model_ids": [
          "google/gemma-4-26b-a4b-it"
        ],
        "transcript_source_artifacts": [
          "run_overnight_cheapest_20260710"
        ],
        "actual_cost_usd": 5.941408519999991,
        "actual_billable_api_calls": 2295
      },
      {
        "artifact_id": "20260710_163315",
        "file": "per_run.jsonl",
        "sha256": "3eba713ba5cc7646df21a7320092f7c1c54906101b73a05178f804ca47ed8a0a",
        "row_count": 63,
        "model_ids": [
          "qwen/qwen3.6-35b-a3b"
        ],
        "transcript_source_artifacts": [
          "run_overnight_cheapest_20260710",
          "run_recovery_qwen_20260710"
        ],
        "actual_cost_usd": 6.424810899999993,
        "actual_billable_api_calls": 2286
      }
    ]
  },
  "models": [
    {
      "model": "Claude Opus 4.8",
      "model_id": "anthropic/claude-opus-4.8",
      "file": "anthropic-claude-opus-4-8.json",
      "sha256": "86275cb442fad4a41c3a388c8d9d8cde64ff563b3ca5cb7ce0d5779799a6afbf",
      "bytes": 2525815,
      "row_count": 63,
      "mode_result_count": 3150
    },
    {
      "model": "Gemma 4 26B A4B",
      "model_id": "google/gemma-4-26b-a4b-it",
      "file": "google-gemma-4-26b-a4b-it.json",
      "sha256": "9ac167a7b86948c5333f108a6dc9318fddefae68a8116c0429c6a831b871370f",
      "bytes": 2504332,
      "row_count": 63,
      "mode_result_count": 3150
    },
    {
      "model": "Gemma 4 31B",
      "model_id": "google/gemma-4-31b-it",
      "file": "google-gemma-4-31b-it.json",
      "sha256": "a3d5b0fb940db41fa386c174375458abe2dd1abb74f406f0c7874008c234751c",
      "bytes": 2508911,
      "row_count": 63,
      "mode_result_count": 3150
    },
    {
      "model": "Qwen 3.6 35B",
      "model_id": "qwen/qwen3.6-35b-a3b",
      "file": "qwen-qwen3-6-35b-a3b.json",
      "sha256": "3901db2ae67076f7afa4efe99791f086c6175095fc0329a529f7178394869423",
      "bytes": 2510379,
      "row_count": 63,
      "mode_result_count": 3150
    }
  ]
}