{
  "run_id": "20260929T174533_gpt-6.1-sol_c86c",
  "model": "gpt-6.1-sol",
  "provider": "OpenAI",
  "effort": "",
  "timestamp": "2026-09-29T17:45:33.054370+00:00",
  "task_count": 6,
  "failed_count": 0,
  "total_time_seconds": 1901.7,
  "time_complete": true,
  "usage_complete": true,
  "previous_attempts": 0,
  "previous_usage": [],
  "prompt_tokens": 38113,
  "output_tokens": 62495,
  "output_tokens_basis": "exact",
  "reasoning_tokens": null,
  "tokens_per_second": 32.86,
  "cost_usd": null,
  "cost_basis": "",
  "cost_complete": false,
  "allowance": null,
  "tasks": [
    {
      "task_id": "aero-lab",
      "task_name": "Aero Lab",
      "task_number": 1,
      "wildcard": false,
      "failed": false,
      "attempts": 1,
      "score": 4.05,
      "time_seconds": 248.32042091700714,
      "ttft_ms": null,
      "prompt_tokens": 6187,
      "output_tokens": 8150,
      "output_tokens_basis": "exact",
      "reasoning_tokens": null,
      "tokens_per_second": 32.82049849103577,
      "cost_usd": null,
      "cost_basis": "",
      "functional": 3.5,
      "visual": 4.0,
      "code": 5.0,
      "ux": 4.0,
      "innovation": null,
      "required_pass_rate": 0.56,
      "judge": "llm-vision",
      "scored_by": "claude-code/claude-opus-5-5 (quality-1 rubric)",
      "prompt_fingerprint": "6c713ca03954f51e",
      "scoring_fingerprint": "59929116d9ab5b1e",
      "rubric": "quality-1",
      "evaluation_history": [
        {
          "score": 5.8,
          "rubric": "legacy",
          "scored_by": "codex-cli/gpt-6.1-sol"
        }
      ],
      "score_note": "",
      "finish_reason": "stop",
      "continuations": 0,
      "html": "01_aero-lab.html",
      "screenshot": "01_aero-lab_screenshot.jpg",
      "frames": "01_aero-lab_frames.jpg",
      "checks": [],
      "lines_changed": null,
      "repair": null
    },
    {
      "task_id": "jet-canyon",
      "task_name": "Jet Canyon",
      "task_number": 2,
      "wildcard": false,
      "failed": false,
      "attempts": 1,
      "score": 7.1,
      "time_seconds": 262.8356817919994,
      "ttft_ms": null,
      "prompt_tokens": 6203,
      "output_tokens": 8609,
      "output_tokens_basis": "exact",
      "reasoning_tokens": null,
      "tokens_per_second": 32.754304671665224,
      "cost_usd": null,
      "cost_basis": "",
      "functional": 8.0,
      "visual": 8.0,
      "code": 6.0,
      "ux": 5.5,
      "innovation": null,
      "required_pass_rate": 0.67,
      "judge": "llm-vision",
      "scored_by": "claude-code/claude-opus-5-5 (quality-1 rubric)",
      "prompt_fingerprint": "021db4866371ec15",
      "scoring_fingerprint": "59929116d9ab5b1e",
      "rubric": "quality-1",
      "evaluation_history": [
        {
          "score": 5.8,
          "rubric": "legacy",
          "scored_by": "codex-cli/gpt-6.1-sol"
        }
      ],
      "score_note": "",
      "finish_reason": "stop",
      "continuations": 0,
      "html": "02_jet-canyon.html",
      "screenshot": "02_jet-canyon_screenshot.jpg",
      "frames": "02_jet-canyon_frames.jpg",
      "checks": [],
      "lines_changed": null,
      "repair": null
    },
    {
      "task_id": "blast-island",
      "task_name": "Blast Island",
      "task_number": 3,
      "wildcard": false,
      "failed": false,
      "attempts": 1,
      "score": 6.45,
      "time_seconds": 324.12096545798704,
      "ttft_ms": null,
      "prompt_tokens": 6534,
      "output_tokens": 10667,
      "output_tokens_basis": "exact",
      "reasoning_tokens": null,
      "tokens_per_second": 32.91055234556455,
      "cost_usd": null,
      "cost_basis": "",
      "functional": 7.0,
      "visual": 7.5,
      "code": 5.5,
      "ux": 5.0,
      "innovation": null,
      "required_pass_rate": 0.44,
      "judge": "llm-vision",
      "scored_by": "claude-code/claude-opus-5-5 (quality-1 rubric)",
      "prompt_fingerprint": "a282381696456fc9",
      "scoring_fingerprint": "59929116d9ab5b1e",
      "rubric": "quality-1",
      "evaluation_history": [
        {
          "score": 6.05,
          "rubric": "legacy",
          "scored_by": "codex-cli/gpt-6.1-sol"
        }
      ],
      "score_note": "",
      "finish_reason": "stop",
      "continuations": 0,
      "html": "03_blast-island.html",
      "screenshot": "03_blast-island_screenshot.jpg",
      "frames": "03_blast-island_frames.jpg",
      "checks": [],
      "lines_changed": null,
      "repair": null
    },
    {
      "task_id": "launch-day",
      "task_name": "Launch Day",
      "task_number": 4,
      "wildcard": false,
      "failed": false,
      "attempts": 1,
      "score": 6.8,
      "time_seconds": 348.0749292499968,
      "ttft_ms": null,
      "prompt_tokens": 7231,
      "output_tokens": 11400,
      "output_tokens_basis": "exact",
      "reasoning_tokens": null,
      "tokens_per_second": 32.75156882043697,
      "cost_usd": null,
      "cost_basis": "",
      "functional": 7.5,
      "visual": 7.5,
      "code": 6.0,
      "ux": 5.5,
      "innovation": null,
      "required_pass_rate": 0.4,
      "judge": "llm-vision",
      "scored_by": "claude-code/claude-opus-5-5 (quality-1 rubric)",
      "prompt_fingerprint": "2787714823d7edc7",
      "scoring_fingerprint": "59929116d9ab5b1e",
      "rubric": "quality-1",
      "evaluation_history": [
        {
          "score": 5.35,
          "rubric": "legacy",
          "scored_by": "codex-cli/gpt-6.1-sol"
        }
      ],
      "score_note": "",
      "finish_reason": "stop",
      "continuations": 0,
      "html": "04_launch-day.html",
      "screenshot": "04_launch-day_screenshot.jpg",
      "frames": "04_launch-day_frames.jpg",
      "checks": [],
      "lines_changed": null,
      "repair": null
    },
    {
      "task_id": "camera-shop",
      "task_name": "Alex Optics",
      "task_number": 5,
      "wildcard": false,
      "failed": false,
      "attempts": 1,
      "score": 6.85,
      "time_seconds": 521.8647030409775,
      "ttft_ms": null,
      "prompt_tokens": 6507,
      "output_tokens": 17273,
      "output_tokens_basis": "exact",
      "reasoning_tokens": null,
      "tokens_per_second": 33.09861713073877,
      "cost_usd": null,
      "cost_basis": "",
      "functional": 7.0,
      "visual": 7.5,
      "code": 6.5,
      "ux": 6.0,
      "innovation": null,
      "required_pass_rate": 0.56,
      "judge": "llm-vision",
      "scored_by": "claude-code/claude-opus-5-5 (quality-1 rubric)",
      "prompt_fingerprint": "e519ec9dadc0a186",
      "scoring_fingerprint": "59929116d9ab5b1e",
      "rubric": "quality-1",
      "evaluation_history": [
        {
          "score": 5.65,
          "rubric": "legacy",
          "scored_by": "codex-cli/gpt-6.1-sol"
        }
      ],
      "score_note": "",
      "finish_reason": "stop",
      "continuations": 0,
      "html": "05_camera-shop.html",
      "screenshot": "05_camera-shop_screenshot.jpg",
      "frames": "05_camera-shop_frames.jpg",
      "checks": [],
      "lines_changed": null,
      "repair": null
    },
    {
      "task_id": "wild-wind-loom-portraits",
      "task_name": "Wind Loom Portraits",
      "task_number": 6,
      "wildcard": true,
      "failed": false,
      "attempts": 1,
      "score": 5.85,
      "time_seconds": 196.47929566598032,
      "ttft_ms": null,
      "prompt_tokens": 5451,
      "output_tokens": 6396,
      "output_tokens_basis": "exact",
      "reasoning_tokens": null,
      "tokens_per_second": 32.553048291018705,
      "cost_usd": null,
      "cost_basis": "",
      "functional": 5.5,
      "visual": 7.0,
      "code": 5.5,
      "ux": 5.0,
      "innovation": null,
      "required_pass_rate": 0.71,
      "judge": "llm-vision",
      "scored_by": "claude-code/claude-opus-5-5 (quality-1 rubric)",
      "prompt_fingerprint": "408e6b279c62918a",
      "scoring_fingerprint": "59929116d9ab5b1e",
      "rubric": "quality-1",
      "evaluation_history": [
        {
          "score": 7.75,
          "rubric": "legacy",
          "scored_by": "codex-cli/gpt-6.1-sol"
        }
      ],
      "score_note": "",
      "finish_reason": "stop",
      "continuations": 0,
      "html": "06_wild-wind-loom-portraits.html",
      "screenshot": "06_wild-wind-loom-portraits_screenshot.jpg",
      "frames": "06_wild-wind-loom-portraits_frames.jpg",
      "checks": [],
      "lines_changed": null,
      "repair": null
    }
  ],
  "score": 6.25,
  "functional": 6.6,
  "visual": 6.9,
  "code": 5.8,
  "ux": 5.2,
  "innovation": 0.0,
  "required_pass_rate": 0.52,
  "wildcard_average": 5.85,
  "ranked": true,
  "folder": "20260929t174533_gpt-6.1-sol_c86c",
  "ran_with": "codex-cli",
  "evidence": {
    "kind": "benchmark",
    "model_details": {
      "display_name": "gpt-6.1-sol",
      "parameters_b": null,
      "active_parameters_b": null,
      "quantization": "",
      "hardware": "",
      "location": "cloud"
    },
    "official_score": null,
    "edition": null,
    "scoring_revision": "",
    "track": "",
    "track_label": "Unknown protocol",
    "coverage": {
      "scored": 0,
      "total": 0
    },
    "evidence_status": "legacy"
  },
  "experiment": null,
  "run_average": 6.25,
  "official_score": null,
  "quality": {
    "standard": "quality-1",
    "label": "Quality v1",
    "score": 6.25,
    "coverage": {
      "scored": 5,
      "total": 5
    },
    "dimensions": {
      "functional": 6.6,
      "visual": 6.9,
      "code": 5.8,
      "ux": 5.2
    },
    "scorers": [
      "claude-code/claude-opus-5-5"
    ]
  }
}