{
  "pass_mode": "pass5",
  "version": "rcb-v1.5-pass5",
  "generated_at": "2026-06-29",
  "source": "ResearchClawBench-OtherRuns/rcb_pass5_scores",
  "tasks": [
    "Astronomy_000",
    "Astronomy_001",
    "Astronomy_002",
    "Astronomy_003",
    "Chemistry_000",
    "Chemistry_001",
    "Chemistry_002",
    "Chemistry_003",
    "Earth_000",
    "Earth_001",
    "Earth_002",
    "Earth_003",
    "Energy_000",
    "Energy_001",
    "Energy_002",
    "Energy_003",
    "Information_000",
    "Information_001",
    "Information_002",
    "Information_003",
    "Life_000",
    "Life_001",
    "Life_002",
    "Life_003",
    "Material_000",
    "Material_001",
    "Material_002",
    "Material_003",
    "Math_000",
    "Math_001",
    "Math_002",
    "Math_003",
    "Neuroscience_000",
    "Neuroscience_001",
    "Neuroscience_002",
    "Neuroscience_003",
    "Physics_000",
    "Physics_001",
    "Physics_002",
    "Physics_003"
  ],
  "agents": [
    "ResearchHarness (Claude-Opus-4.8)",
    "ResearchHarness (Gemini-3.5-Flash)",
    "ResearchHarness (Kimi-K2.6)",
    "ResearchHarness (GPT-5.5)",
    "ResearchHarness (MiMo-V2.5)",
    "ResearchHarness (GLM-5.1)"
  ],
  "scores": {
    "ResearchHarness (Claude-Opus-4.8)": {
      "Astronomy_000": {
        "score": 26.6,
        "mean": 23.48,
        "std": 1.8,
        "min": 21.1,
        "max": 26.6,
        "runs": [
          26.6,
          23.6,
          23.5,
          22.6,
          21.1
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Astronomy_001": {
        "score": 40.6,
        "mean": 30.68,
        "std": 7.53,
        "min": 19.8,
        "max": 40.6,
        "runs": [
          40.6,
          36.2,
          32.0,
          24.8,
          19.8
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Astronomy_002": {
        "score": 29.9,
        "mean": 25.5,
        "std": 4.69,
        "min": 18.1,
        "max": 29.9,
        "runs": [
          29.9,
          29.5,
          28.1,
          21.9,
          18.1
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Astronomy_003": {
        "score": 48.6,
        "mean": 45.78,
        "std": 1.77,
        "min": 43.9,
        "max": 48.6,
        "runs": [
          48.6,
          46.8,
          45.6,
          44.0,
          43.9
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Chemistry_000": {
        "score": 34.0,
        "mean": 25.35,
        "std": 6.25,
        "min": 14.95,
        "max": 34.0,
        "runs": [
          34.0,
          27.75,
          26.85,
          23.2,
          14.95
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Chemistry_001": {
        "score": 9.65,
        "mean": 9.12,
        "std": 0.37,
        "min": 8.65,
        "max": 9.65,
        "runs": [
          9.65,
          9.3,
          9.25,
          8.75,
          8.65
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Chemistry_002": {
        "score": 18.2,
        "mean": 9.79,
        "std": 6.19,
        "min": 1.75,
        "max": 18.2,
        "runs": [
          18.2,
          15.7,
          7.5,
          5.8,
          1.75
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Chemistry_003": {
        "score": 30.6,
        "mean": 27.76,
        "std": 1.76,
        "min": 25.8,
        "max": 30.6,
        "runs": [
          30.6,
          28.5,
          27.9,
          26.0,
          25.8
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Earth_000": {
        "score": 23.6,
        "mean": 17.52,
        "std": 4.36,
        "min": 12.8,
        "max": 23.6,
        "runs": [
          23.6,
          21.3,
          17.0,
          12.9,
          12.8
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Earth_001": {
        "score": 38.68,
        "mean": 32.7,
        "std": 4.69,
        "min": 24.54,
        "max": 38.68,
        "runs": [
          38.68,
          34.72,
          34.1,
          31.45,
          24.54
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Earth_002": {
        "score": 34.1,
        "mean": 28.48,
        "std": 7.56,
        "min": 14.4,
        "max": 34.1,
        "runs": [
          34.1,
          34.0,
          33.2,
          26.7,
          14.4
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Earth_003": {
        "score": 8.8,
        "mean": 5.4,
        "std": 3.85,
        "min": 0.4,
        "max": 8.8,
        "runs": [
          8.8,
          8.8,
          8.0,
          1.0,
          0.4
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Energy_000": {
        "score": 25.0,
        "mean": 15.2,
        "std": 6.98,
        "min": 6.0,
        "max": 25.0,
        "runs": [
          25.0,
          19.9,
          16.2,
          8.9,
          6.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Energy_001": {
        "score": 31.7,
        "mean": 23.82,
        "std": 5.85,
        "min": 15.0,
        "max": 31.7,
        "runs": [
          31.7,
          26.6,
          26.2,
          19.6,
          15.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Energy_002": {
        "score": 39.8,
        "mean": 24.04,
        "std": 14.32,
        "min": 0.0,
        "max": 39.8,
        "runs": [
          39.8,
          36.05,
          27.05,
          17.3,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Energy_003": {
        "score": 25.9,
        "mean": 15.0,
        "std": 7.88,
        "min": 7.5,
        "max": 25.9,
        "runs": [
          25.9,
          23.0,
          11.1,
          7.5,
          7.5
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Information_000": {
        "score": 58.6,
        "mean": 24.46,
        "std": 18.74,
        "min": 3.2,
        "max": 58.6,
        "runs": [
          58.6,
          25.5,
          22.0,
          13.0,
          3.2
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Information_001": {
        "score": 19.6,
        "mean": 7.0,
        "std": 6.56,
        "min": 1.0,
        "max": 19.6,
        "runs": [
          19.6,
          6.4,
          5.0,
          3.0,
          1.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Information_002": {
        "score": 38.4,
        "mean": 28.56,
        "std": 6.72,
        "min": 18.9,
        "max": 38.4,
        "runs": [
          38.4,
          32.6,
          28.7,
          24.2,
          18.9
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Information_003": {
        "score": 21.85,
        "mean": 14.93,
        "std": 3.79,
        "min": 10.3,
        "max": 21.85,
        "runs": [
          21.85,
          14.85,
          14.0,
          13.65,
          10.3
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Life_000": {
        "score": 12.55,
        "mean": 10.27,
        "std": 2.14,
        "min": 7.3,
        "max": 12.55,
        "runs": [
          12.55,
          12.3,
          11.0,
          8.2,
          7.3
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Life_001": {
        "score": 15.35,
        "mean": 9.7,
        "std": 5.29,
        "min": 0.0,
        "max": 15.35,
        "runs": [
          15.35,
          12.3,
          12.2,
          8.65,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Life_002": {
        "score": 14.0,
        "mean": 10.16,
        "std": 2.42,
        "min": 7.8,
        "max": 14.0,
        "runs": [
          14.0,
          12.0,
          8.8,
          8.2,
          7.8
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Life_003": {
        "score": 41.3,
        "mean": 33.24,
        "std": 5.53,
        "min": 25.6,
        "max": 41.3,
        "runs": [
          41.3,
          37.3,
          32.2,
          29.8,
          25.6
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Material_000": {
        "score": 36.8,
        "mean": 26.44,
        "std": 5.55,
        "min": 20.1,
        "max": 36.8,
        "runs": [
          36.8,
          25.9,
          25.0,
          24.4,
          20.1
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Material_001": {
        "score": 31.05,
        "mean": 25.36,
        "std": 5.3,
        "min": 16.35,
        "max": 31.05,
        "runs": [
          31.05,
          28.5,
          28.35,
          22.55,
          16.35
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Material_002": {
        "score": 43.71,
        "mean": 36.31,
        "std": 4.97,
        "min": 29.52,
        "max": 43.71,
        "runs": [
          43.71,
          39.73,
          35.44,
          33.14,
          29.52
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Material_003": {
        "score": 26.0,
        "mean": 21.02,
        "std": 3.78,
        "min": 15.15,
        "max": 26.0,
        "runs": [
          26.0,
          23.75,
          21.3,
          18.9,
          15.15
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Math_000": {
        "score": 32.05,
        "mean": 28.52,
        "std": 2.96,
        "min": 24.9,
        "max": 32.05,
        "runs": [
          32.05,
          31.9,
          27.75,
          26.0,
          24.9
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Math_001": {
        "score": 46.4,
        "mean": 34.6,
        "std": 6.86,
        "min": 25.0,
        "max": 46.4,
        "runs": [
          46.4,
          34.9,
          33.9,
          32.8,
          25.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Math_002": {
        "score": 26.5,
        "mean": 14.24,
        "std": 9.81,
        "min": 0.0,
        "max": 26.5,
        "runs": [
          26.5,
          24.1,
          10.4,
          10.2,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Math_003": {
        "score": 15.75,
        "mean": 5.63,
        "std": 6.98,
        "min": 0.0,
        "max": 15.75,
        "runs": [
          15.75,
          12.4,
          0.0,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Neuroscience_000": {
        "score": 14.2,
        "mean": 12.56,
        "std": 1.76,
        "min": 9.4,
        "max": 14.2,
        "runs": [
          14.2,
          14.0,
          13.2,
          12.0,
          9.4
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Neuroscience_001": {
        "score": 25.95,
        "mean": 22.09,
        "std": 3.37,
        "min": 16.0,
        "max": 25.95,
        "runs": [
          25.95,
          23.8,
          23.25,
          21.45,
          16.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Neuroscience_002": {
        "score": 6.25,
        "mean": 3.4,
        "std": 1.79,
        "min": 1.25,
        "max": 6.25,
        "runs": [
          6.25,
          4.5,
          3.0,
          2.0,
          1.25
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Neuroscience_003": {
        "score": 20.75,
        "mean": 15.5,
        "std": 3.72,
        "min": 11.05,
        "max": 20.75,
        "runs": [
          20.75,
          17.5,
          16.75,
          11.45,
          11.05
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Physics_000": {
        "score": 37.8,
        "mean": 29.66,
        "std": 5.83,
        "min": 24.8,
        "max": 37.8,
        "runs": [
          37.8,
          35.7,
          25.0,
          25.0,
          24.8
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Physics_001": {
        "score": 34.2,
        "mean": 32.26,
        "std": 1.56,
        "min": 29.45,
        "max": 34.2,
        "runs": [
          34.2,
          32.95,
          32.4,
          32.3,
          29.45
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Physics_002": {
        "score": 41.2,
        "mean": 36.42,
        "std": 2.56,
        "min": 34.1,
        "max": 41.2,
        "runs": [
          41.2,
          36.3,
          36.2,
          34.3,
          34.1
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Physics_003": {
        "score": 64.4,
        "mean": 50.82,
        "std": 6.9,
        "min": 46.0,
        "max": 64.4,
        "runs": [
          64.4,
          49.4,
          47.9,
          46.4,
          46.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Claude-Opus-4.8",
        "model_display": "Claude-Opus-4.8",
        "details_exported": false,
        "pass_mode": "pass5"
      }
    },
    "ResearchHarness (Gemini-3.5-Flash)": {
      "Astronomy_000": {
        "score": 32.6,
        "mean": 19.52,
        "std": 9.29,
        "min": 8.0,
        "max": 32.6,
        "runs": [
          32.6,
          28.2,
          14.6,
          14.2,
          8.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Astronomy_001": {
        "score": 37.6,
        "mean": 24.96,
        "std": 11.48,
        "min": 7.2,
        "max": 37.6,
        "runs": [
          37.6,
          36.8,
          24.8,
          18.4,
          7.2
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Astronomy_002": {
        "score": 24.0,
        "mean": 4.8,
        "std": 9.6,
        "min": 0.0,
        "max": 24.0,
        "runs": [
          24.0,
          0.0,
          0.0,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Astronomy_003": {
        "score": 46.6,
        "mean": 43.86,
        "std": 1.95,
        "min": 40.8,
        "max": 46.6,
        "runs": [
          46.6,
          44.7,
          44.4,
          42.8,
          40.8
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Chemistry_000": {
        "score": 19.55,
        "mean": 18.01,
        "std": 1.24,
        "min": 16.3,
        "max": 19.55,
        "runs": [
          19.55,
          18.9,
          18.45,
          16.85,
          16.3
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Chemistry_001": {
        "score": 8.55,
        "mean": 4.18,
        "std": 2.6,
        "min": 1.25,
        "max": 8.55,
        "runs": [
          8.55,
          5.0,
          4.25,
          1.85,
          1.25
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Chemistry_002": {
        "score": 6.0,
        "mean": 1.4,
        "std": 2.33,
        "min": 0.0,
        "max": 6.0,
        "runs": [
          6.0,
          1.0,
          0.0,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Chemistry_003": {
        "score": 28.2,
        "mean": 12.12,
        "std": 9.21,
        "min": 0.0,
        "max": 28.2,
        "runs": [
          28.2,
          13.5,
          10.5,
          8.4,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Earth_000": {
        "score": 23.6,
        "mean": 19.42,
        "std": 4.21,
        "min": 11.7,
        "max": 23.6,
        "runs": [
          23.6,
          22.8,
          19.6,
          19.4,
          11.7
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Earth_001": {
        "score": 39.95,
        "mean": 35.46,
        "std": 3.14,
        "min": 31.18,
        "max": 39.95,
        "runs": [
          39.95,
          37.04,
          36.38,
          32.75,
          31.18
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Earth_002": {
        "score": 31.7,
        "mean": 26.54,
        "std": 3.56,
        "min": 22.9,
        "max": 31.7,
        "runs": [
          31.7,
          29.9,
          24.5,
          23.7,
          22.9
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Earth_003": {
        "score": 13.0,
        "mean": 8.06,
        "std": 3.66,
        "min": 2.5,
        "max": 13.0,
        "runs": [
          13.0,
          11.0,
          7.3,
          6.5,
          2.5
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Energy_000": {
        "score": 20.5,
        "mean": 16.14,
        "std": 4.24,
        "min": 8.0,
        "max": 20.5,
        "runs": [
          20.5,
          17.5,
          17.5,
          17.2,
          8.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Energy_001": {
        "score": 24.5,
        "mean": 18.54,
        "std": 3.44,
        "min": 15.1,
        "max": 24.5,
        "runs": [
          24.5,
          20.0,
          17.5,
          15.6,
          15.1
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Energy_002": {
        "score": 38.65,
        "mean": 36.76,
        "std": 1.73,
        "min": 34.05,
        "max": 38.65,
        "runs": [
          38.65,
          38.55,
          36.75,
          35.8,
          34.05
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Energy_003": {
        "score": 44.0,
        "mean": 23.44,
        "std": 11.71,
        "min": 7.5,
        "max": 44.0,
        "runs": [
          44.0,
          22.7,
          22.0,
          21.0,
          7.5
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Information_000": {
        "score": 41.4,
        "mean": 24.98,
        "std": 8.31,
        "min": 18.5,
        "max": 41.4,
        "runs": [
          41.4,
          22.5,
          21.3,
          21.2,
          18.5
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Information_001": {
        "score": 17.6,
        "mean": 7.8,
        "std": 6.39,
        "min": 1.6,
        "max": 17.6,
        "runs": [
          17.6,
          13.2,
          3.6,
          3.0,
          1.6
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Information_002": {
        "score": 32.6,
        "mean": 27.94,
        "std": 4.18,
        "min": 21.8,
        "max": 32.6,
        "runs": [
          32.6,
          30.9,
          30.2,
          24.2,
          21.8
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Information_003": {
        "score": 17.6,
        "mean": 11.56,
        "std": 5.41,
        "min": 4.6,
        "max": 17.6,
        "runs": [
          17.6,
          16.4,
          13.5,
          5.7,
          4.6
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Life_000": {
        "score": 9.3,
        "mean": 7.84,
        "std": 1.35,
        "min": 6.25,
        "max": 9.3,
        "runs": [
          9.3,
          9.25,
          8.1,
          6.3,
          6.25
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Life_001": {
        "score": 29.0,
        "mean": 12.69,
        "std": 8.96,
        "min": 3.75,
        "max": 29.0,
        "runs": [
          29.0,
          13.95,
          11.25,
          5.5,
          3.75
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Life_002": {
        "score": 24.0,
        "mean": 9.12,
        "std": 7.54,
        "min": 3.4,
        "max": 24.0,
        "runs": [
          24.0,
          7.0,
          6.2,
          5.0,
          3.4
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Life_003": {
        "score": 41.6,
        "mean": 33.24,
        "std": 4.44,
        "min": 29.2,
        "max": 41.6,
        "runs": [
          41.6,
          32.8,
          32.8,
          29.8,
          29.2
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Material_000": {
        "score": 24.1,
        "mean": 20.5,
        "std": 3.47,
        "min": 14.4,
        "max": 24.1,
        "runs": [
          24.1,
          23.6,
          20.4,
          20.0,
          14.4
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Material_001": {
        "score": 13.05,
        "mean": 9.96,
        "std": 2.34,
        "min": 7.0,
        "max": 13.05,
        "runs": [
          13.05,
          12.25,
          9.45,
          8.05,
          7.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Material_002": {
        "score": 42.03,
        "mean": 38.42,
        "std": 2.82,
        "min": 34.47,
        "max": 42.03,
        "runs": [
          42.03,
          41.07,
          38.08,
          36.43,
          34.47
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Material_003": {
        "score": 18.5,
        "mean": 16.48,
        "std": 1.46,
        "min": 14.1,
        "max": 18.5,
        "runs": [
          18.5,
          17.25,
          16.65,
          15.9,
          14.1
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Math_000": {
        "score": 45.0,
        "mean": 25.63,
        "std": 10.55,
        "min": 13.55,
        "max": 45.0,
        "runs": [
          45.0,
          26.55,
          22.05,
          21.0,
          13.55
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Math_001": {
        "score": 39.2,
        "mean": 26.56,
        "std": 6.78,
        "min": 19.0,
        "max": 39.2,
        "runs": [
          39.2,
          26.4,
          24.6,
          23.6,
          19.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Math_002": {
        "score": 18.1,
        "mean": 11.24,
        "std": 3.98,
        "min": 6.9,
        "max": 18.1,
        "runs": [
          18.1,
          12.4,
          11.0,
          7.8,
          6.9
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Math_003": {
        "score": 13.75,
        "mean": 6.19,
        "std": 5.46,
        "min": 0.0,
        "max": 13.75,
        "runs": [
          13.75,
          10.0,
          7.2,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Neuroscience_000": {
        "score": 18.4,
        "mean": 10.4,
        "std": 4.58,
        "min": 5.0,
        "max": 18.4,
        "runs": [
          18.4,
          12.0,
          8.6,
          8.0,
          5.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Neuroscience_001": {
        "score": 9.25,
        "mean": 6.46,
        "std": 3.14,
        "min": 0.75,
        "max": 9.25,
        "runs": [
          9.25,
          9.25,
          7.3,
          5.75,
          0.75
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Neuroscience_002": {
        "score": 4.85,
        "mean": 1.57,
        "std": 1.66,
        "min": 0.4,
        "max": 4.85,
        "runs": [
          4.85,
          1.0,
          1.0,
          0.6,
          0.4
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Neuroscience_003": {
        "score": 15.0,
        "mean": 12.16,
        "std": 1.91,
        "min": 9.45,
        "max": 15.0,
        "runs": [
          15.0,
          12.9,
          12.7,
          10.75,
          9.45
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Physics_000": {
        "score": 35.0,
        "mean": 21.04,
        "std": 7.3,
        "min": 15.2,
        "max": 35.0,
        "runs": [
          35.0,
          21.0,
          18.6,
          15.4,
          15.2
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Physics_001": {
        "score": 37.65,
        "mean": 23.86,
        "std": 12.85,
        "min": 0.0,
        "max": 37.65,
        "runs": [
          37.65,
          31.15,
          27.45,
          23.05,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Physics_002": {
        "score": 40.7,
        "mean": 27.4,
        "std": 14.4,
        "min": 0.0,
        "max": 40.7,
        "runs": [
          40.7,
          37.4,
          30.25,
          28.65,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Physics_003": {
        "score": 47.8,
        "mean": 39.94,
        "std": 8.04,
        "min": 25.6,
        "max": 47.8,
        "runs": [
          47.8,
          46.5,
          42.4,
          37.4,
          25.6
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Gemini-3.5-Flash",
        "model_display": "Gemini-3.5-Flash",
        "details_exported": false,
        "pass_mode": "pass5"
      }
    },
    "ResearchHarness (Kimi-K2.6)": {
      "Astronomy_000": {
        "score": 22.6,
        "mean": 12.72,
        "std": 7.54,
        "min": 0.0,
        "max": 22.6,
        "runs": [
          22.6,
          16.5,
          14.5,
          10.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Astronomy_001": {
        "score": 36.0,
        "mean": 17.28,
        "std": 11.52,
        "min": 0.0,
        "max": 36.0,
        "runs": [
          36.0,
          19.2,
          17.2,
          14.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Astronomy_002": {
        "score": 28.0,
        "mean": 9.9,
        "std": 12.3,
        "min": 0.0,
        "max": 28.0,
        "runs": [
          28.0,
          21.5,
          0.0,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Astronomy_003": {
        "score": 50.5,
        "mean": 47.9,
        "std": 1.41,
        "min": 46.4,
        "max": 50.5,
        "runs": [
          50.5,
          48.0,
          47.6,
          47.0,
          46.4
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Chemistry_000": {
        "score": 18.55,
        "mean": 6.31,
        "std": 7.92,
        "min": 0.0,
        "max": 18.55,
        "runs": [
          18.55,
          13.0,
          0.0,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Chemistry_001": {
        "score": 10.85,
        "mean": 6.57,
        "std": 3.13,
        "min": 1.25,
        "max": 10.85,
        "runs": [
          10.85,
          8.0,
          6.75,
          6.0,
          1.25
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Chemistry_002": {
        "score": 1.75,
        "mean": 0.98,
        "std": 0.44,
        "min": 0.4,
        "max": 1.75,
        "runs": [
          1.75,
          1.0,
          1.0,
          0.75,
          0.4
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Chemistry_003": {
        "score": 15.9,
        "mean": 7.98,
        "std": 5.8,
        "min": 0.0,
        "max": 15.9,
        "runs": [
          15.9,
          12.0,
          9.0,
          3.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Earth_000": {
        "score": 23.8,
        "mean": 20.7,
        "std": 2.09,
        "min": 17.8,
        "max": 23.8,
        "runs": [
          23.8,
          22.2,
          20.0,
          19.7,
          17.8
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Earth_001": {
        "score": 44.23,
        "mean": 35.04,
        "std": 8.63,
        "min": 22.28,
        "max": 44.23,
        "runs": [
          44.23,
          42.87,
          38.15,
          27.68,
          22.28
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Earth_002": {
        "score": 36.1,
        "mean": 29.82,
        "std": 4.31,
        "min": 25.1,
        "max": 36.1,
        "runs": [
          36.1,
          33.6,
          28.3,
          26.0,
          25.1
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Earth_003": {
        "score": 5.7,
        "mean": 1.9,
        "std": 1.93,
        "min": 0.4,
        "max": 5.7,
        "runs": [
          5.7,
          1.5,
          1.0,
          0.9,
          0.4
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Energy_000": {
        "score": 22.0,
        "mean": 15.06,
        "std": 5.35,
        "min": 7.5,
        "max": 22.0,
        "runs": [
          22.0,
          19.0,
          16.3,
          10.5,
          7.5
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Energy_001": {
        "score": 27.8,
        "mean": 12.72,
        "std": 11.02,
        "min": 0.0,
        "max": 27.8,
        "runs": [
          27.8,
          18.9,
          16.9,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Energy_002": {
        "score": 39.2,
        "mean": 34.27,
        "std": 3.77,
        "min": 28.0,
        "max": 39.2,
        "runs": [
          39.2,
          36.7,
          34.4,
          33.05,
          28.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Energy_003": {
        "score": 25.0,
        "mean": 12.22,
        "std": 6.72,
        "min": 6.5,
        "max": 25.0,
        "runs": [
          25.0,
          12.6,
          9.5,
          7.5,
          6.5
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Information_000": {
        "score": 41.5,
        "mean": 24.84,
        "std": 12.29,
        "min": 4.8,
        "max": 41.5,
        "runs": [
          41.5,
          31.4,
          27.0,
          19.5,
          4.8
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Information_001": {
        "score": 7.6,
        "mean": 6.08,
        "std": 1.48,
        "min": 3.6,
        "max": 7.6,
        "runs": [
          7.6,
          7.6,
          6.0,
          5.6,
          3.6
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Information_002": {
        "score": 27.8,
        "mean": 14.68,
        "std": 6.7,
        "min": 9.6,
        "max": 27.8,
        "runs": [
          27.8,
          12.9,
          12.9,
          10.2,
          9.6
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Information_003": {
        "score": 10.7,
        "mean": 5.57,
        "std": 3.66,
        "min": 1.35,
        "max": 10.7,
        "runs": [
          10.7,
          9.15,
          3.4,
          3.25,
          1.35
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Life_000": {
        "score": 10.7,
        "mean": 8.51,
        "std": 2.02,
        "min": 5.0,
        "max": 10.7,
        "runs": [
          10.7,
          10.3,
          8.3,
          8.25,
          5.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Life_001": {
        "score": 12.75,
        "mean": 6.34,
        "std": 4.29,
        "min": 0.0,
        "max": 12.75,
        "runs": [
          12.75,
          8.95,
          5.5,
          4.5,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Life_002": {
        "score": 17.4,
        "mean": 8.84,
        "std": 4.41,
        "min": 5.4,
        "max": 17.4,
        "runs": [
          17.4,
          8.6,
          6.8,
          6.0,
          5.4
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Life_003": {
        "score": 42.0,
        "mean": 32.62,
        "std": 6.07,
        "min": 25.0,
        "max": 42.0,
        "runs": [
          42.0,
          37.0,
          29.7,
          29.4,
          25.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Material_000": {
        "score": 26.0,
        "mean": 20.7,
        "std": 3.87,
        "min": 15.9,
        "max": 26.0,
        "runs": [
          26.0,
          23.4,
          21.5,
          16.7,
          15.9
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Material_001": {
        "score": 25.9,
        "mean": 13.97,
        "std": 8.39,
        "min": 3.25,
        "max": 25.9,
        "runs": [
          25.9,
          19.95,
          14.45,
          6.3,
          3.25
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Material_002": {
        "score": 36.09,
        "mean": 27.13,
        "std": 5.48,
        "min": 21.07,
        "max": 36.09,
        "runs": [
          36.09,
          30.48,
          25.1,
          22.89,
          21.07
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Material_003": {
        "score": 20.9,
        "mean": 17.0,
        "std": 4.06,
        "min": 9.65,
        "max": 20.9,
        "runs": [
          20.9,
          19.8,
          18.9,
          15.75,
          9.65
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Math_000": {
        "score": 33.55,
        "mean": 29.2,
        "std": 3.24,
        "min": 24.0,
        "max": 33.55,
        "runs": [
          33.55,
          31.45,
          28.95,
          28.05,
          24.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Math_001": {
        "score": 37.6,
        "mean": 25.54,
        "std": 11.15,
        "min": 11.6,
        "max": 37.6,
        "runs": [
          37.6,
          33.0,
          33.0,
          12.5,
          11.6
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Math_002": {
        "score": 19.4,
        "mean": 12.88,
        "std": 6.78,
        "min": 2.0,
        "max": 19.4,
        "runs": [
          19.4,
          18.6,
          16.4,
          8.0,
          2.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Math_003": {
        "score": 24.75,
        "mean": 15.15,
        "std": 5.4,
        "min": 10.0,
        "max": 24.75,
        "runs": [
          24.75,
          15.75,
          15.25,
          10.0,
          10.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Neuroscience_000": {
        "score": 19.0,
        "mean": 5.48,
        "std": 7.15,
        "min": 0.0,
        "max": 19.0,
        "runs": [
          19.0,
          6.4,
          2.0,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Neuroscience_001": {
        "score": 12.7,
        "mean": 7.93,
        "std": 3.59,
        "min": 2.0,
        "max": 12.7,
        "runs": [
          12.7,
          10.25,
          7.75,
          6.95,
          2.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Neuroscience_002": {
        "score": 0.75,
        "mean": 0.25,
        "std": 0.32,
        "min": 0.0,
        "max": 0.75,
        "runs": [
          0.75,
          0.5,
          0.0,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Neuroscience_003": {
        "score": 16.2,
        "mean": 12.97,
        "std": 4.51,
        "min": 4.2,
        "max": 16.2,
        "runs": [
          16.2,
          15.8,
          15.5,
          13.15,
          4.2
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Physics_000": {
        "score": 35.6,
        "mean": 28.28,
        "std": 6.36,
        "min": 19.6,
        "max": 35.6,
        "runs": [
          35.6,
          35.4,
          27.1,
          23.7,
          19.6
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Physics_001": {
        "score": 31.3,
        "mean": 26.71,
        "std": 4.6,
        "min": 19.4,
        "max": 31.3,
        "runs": [
          31.3,
          29.95,
          29.65,
          23.25,
          19.4
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Physics_002": {
        "score": 34.4,
        "mean": 27.83,
        "std": 4.12,
        "min": 21.5,
        "max": 34.4,
        "runs": [
          34.4,
          28.7,
          27.6,
          26.95,
          21.5
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Physics_003": {
        "score": 44.8,
        "mean": 33.86,
        "std": 7.62,
        "min": 26.6,
        "max": 44.8,
        "runs": [
          44.8,
          41.2,
          29.6,
          27.1,
          26.6
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "Kimi-K2.6",
        "model_display": "Kimi-K2.6",
        "details_exported": false,
        "pass_mode": "pass5"
      }
    },
    "ResearchHarness (GPT-5.5)": {
      "Astronomy_000": {
        "score": 21.4,
        "mean": 18.04,
        "std": 3.95,
        "min": 10.5,
        "max": 21.4,
        "runs": [
          21.4,
          21.0,
          19.1,
          18.2,
          10.5
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Astronomy_001": {
        "score": 22.4,
        "mean": 16.6,
        "std": 8.43,
        "min": 0.0,
        "max": 22.4,
        "runs": [
          22.4,
          21.6,
          21.0,
          18.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Astronomy_002": {
        "score": 23.5,
        "mean": 20.22,
        "std": 3.83,
        "min": 13.5,
        "max": 23.5,
        "runs": [
          23.5,
          23.5,
          22.1,
          18.5,
          13.5
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Astronomy_003": {
        "score": 50.8,
        "mean": 47.46,
        "std": 2.7,
        "min": 44.1,
        "max": 50.8,
        "runs": [
          50.8,
          50.5,
          46.2,
          45.7,
          44.1
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Chemistry_000": {
        "score": 20.6,
        "mean": 15.08,
        "std": 3.74,
        "min": 10.15,
        "max": 20.6,
        "runs": [
          20.6,
          18.05,
          13.4,
          13.2,
          10.15
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Chemistry_001": {
        "score": 9.0,
        "mean": 4.36,
        "std": 3.62,
        "min": 0.75,
        "max": 9.0,
        "runs": [
          9.0,
          8.5,
          2.3,
          1.25,
          0.75
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Chemistry_002": {
        "score": 3.5,
        "mean": 1.52,
        "std": 1.29,
        "min": 0.0,
        "max": 3.5,
        "runs": [
          3.5,
          2.5,
          1.0,
          0.6,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Chemistry_003": {
        "score": 21.6,
        "mean": 16.74,
        "std": 2.8,
        "min": 12.9,
        "max": 21.6,
        "runs": [
          21.6,
          17.0,
          16.1,
          16.1,
          12.9
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Earth_000": {
        "score": 24.0,
        "mean": 19.78,
        "std": 3.64,
        "min": 13.7,
        "max": 24.0,
        "runs": [
          24.0,
          23.0,
          19.2,
          19.0,
          13.7
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Earth_001": {
        "score": 41.16,
        "mean": 34.9,
        "std": 6.14,
        "min": 27.02,
        "max": 41.16,
        "runs": [
          41.16,
          41.15,
          36.92,
          28.25,
          27.02
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Earth_002": {
        "score": 25.3,
        "mean": 23.52,
        "std": 1.94,
        "min": 20.0,
        "max": 25.3,
        "runs": [
          25.3,
          24.7,
          24.7,
          22.9,
          20.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Earth_003": {
        "score": 1.5,
        "mean": 0.9,
        "std": 0.57,
        "min": 0.0,
        "max": 1.5,
        "runs": [
          1.5,
          1.5,
          0.9,
          0.6,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Energy_000": {
        "score": 18.0,
        "mean": 16.3,
        "std": 1.47,
        "min": 14.4,
        "max": 18.0,
        "runs": [
          18.0,
          17.5,
          16.9,
          14.7,
          14.4
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Energy_001": {
        "score": 18.6,
        "mean": 13.98,
        "std": 2.96,
        "min": 10.7,
        "max": 18.6,
        "runs": [
          18.6,
          15.5,
          14.2,
          10.9,
          10.7
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Energy_002": {
        "score": 40.9,
        "mean": 36.97,
        "std": 2.36,
        "min": 33.95,
        "max": 40.9,
        "runs": [
          40.9,
          37.85,
          36.75,
          35.4,
          33.95
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Energy_003": {
        "score": 23.5,
        "mean": 19.16,
        "std": 3.74,
        "min": 13.5,
        "max": 23.5,
        "runs": [
          23.5,
          22.7,
          19.5,
          16.6,
          13.5
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Information_000": {
        "score": 21.3,
        "mean": 17.36,
        "std": 4.85,
        "min": 8.0,
        "max": 21.3,
        "runs": [
          21.3,
          20.7,
          19.2,
          17.6,
          8.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Information_001": {
        "score": 7.0,
        "mean": 4.32,
        "std": 2.02,
        "min": 1.0,
        "max": 7.0,
        "runs": [
          7.0,
          5.6,
          4.4,
          3.6,
          1.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Information_002": {
        "score": 45.3,
        "mean": 39.8,
        "std": 3.09,
        "min": 37.1,
        "max": 45.3,
        "runs": [
          45.3,
          41.1,
          37.8,
          37.7,
          37.1
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Information_003": {
        "score": 12.4,
        "mean": 9.27,
        "std": 2.34,
        "min": 5.3,
        "max": 12.4,
        "runs": [
          12.4,
          10.3,
          9.75,
          8.6,
          5.3
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Life_000": {
        "score": 11.3,
        "mean": 9.01,
        "std": 1.41,
        "min": 7.3,
        "max": 11.3,
        "runs": [
          11.3,
          9.4,
          9.3,
          7.75,
          7.3
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Life_001": {
        "score": 18.05,
        "mean": 13.76,
        "std": 2.74,
        "min": 9.9,
        "max": 18.05,
        "runs": [
          18.05,
          15.25,
          13.05,
          12.55,
          9.9
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Life_002": {
        "score": 12.8,
        "mean": 10.06,
        "std": 2.49,
        "min": 5.7,
        "max": 12.8,
        "runs": [
          12.8,
          12.0,
          10.5,
          9.3,
          5.7
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Life_003": {
        "score": 56.5,
        "mean": 40.28,
        "std": 9.22,
        "min": 28.2,
        "max": 56.5,
        "runs": [
          56.5,
          41.5,
          37.7,
          37.5,
          28.2
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Material_000": {
        "score": 23.5,
        "mean": 18.64,
        "std": 4.59,
        "min": 9.9,
        "max": 23.5,
        "runs": [
          23.5,
          20.4,
          20.0,
          19.4,
          9.9
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Material_001": {
        "score": 13.0,
        "mean": 8.78,
        "std": 2.26,
        "min": 6.3,
        "max": 13.0,
        "runs": [
          13.0,
          8.75,
          8.05,
          7.8,
          6.3
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Material_002": {
        "score": 25.0,
        "mean": 13.48,
        "std": 7.11,
        "min": 5.1,
        "max": 25.0,
        "runs": [
          25.0,
          17.79,
          11.04,
          8.45,
          5.1
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Material_003": {
        "score": 28.8,
        "mean": 21.9,
        "std": 4.11,
        "min": 16.65,
        "max": 28.8,
        "runs": [
          28.8,
          23.4,
          21.3,
          19.35,
          16.65
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Math_000": {
        "score": 28.0,
        "mean": 22.65,
        "std": 7.51,
        "min": 7.95,
        "max": 28.0,
        "runs": [
          28.0,
          26.95,
          26.95,
          23.4,
          7.95
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Math_001": {
        "score": 46.2,
        "mean": 35.76,
        "std": 7.21,
        "min": 24.7,
        "max": 46.2,
        "runs": [
          46.2,
          40.1,
          35.1,
          32.7,
          24.7
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Math_002": {
        "score": 20.1,
        "mean": 15.72,
        "std": 4.61,
        "min": 7.5,
        "max": 20.1,
        "runs": [
          20.1,
          20.1,
          15.9,
          15.0,
          7.5
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Math_003": {
        "score": 10.0,
        "mean": 8.11,
        "std": 1.29,
        "min": 6.0,
        "max": 10.0,
        "runs": [
          10.0,
          8.4,
          8.4,
          7.75,
          6.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Neuroscience_000": {
        "score": 10.6,
        "mean": 7.64,
        "std": 2.59,
        "min": 3.6,
        "max": 10.6,
        "runs": [
          10.6,
          10.0,
          8.0,
          6.0,
          3.6
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Neuroscience_001": {
        "score": 4.2,
        "mean": 2.33,
        "std": 1.12,
        "min": 1.2,
        "max": 4.2,
        "runs": [
          4.2,
          3.0,
          1.75,
          1.5,
          1.2
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Neuroscience_002": {
        "score": 3.0,
        "mean": 2.0,
        "std": 0.89,
        "min": 1.0,
        "max": 3.0,
        "runs": [
          3.0,
          3.0,
          2.0,
          1.0,
          1.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Neuroscience_003": {
        "score": 11.55,
        "mean": 8.03,
        "std": 2.49,
        "min": 4.5,
        "max": 11.55,
        "runs": [
          11.55,
          9.35,
          8.7,
          6.05,
          4.5
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Physics_000": {
        "score": 36.0,
        "mean": 29.46,
        "std": 5.04,
        "min": 24.0,
        "max": 36.0,
        "runs": [
          36.0,
          35.1,
          26.4,
          25.8,
          24.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Physics_001": {
        "score": 35.55,
        "mean": 30.95,
        "std": 2.87,
        "min": 27.65,
        "max": 35.55,
        "runs": [
          35.55,
          32.05,
          31.35,
          28.15,
          27.65
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Physics_002": {
        "score": 28.25,
        "mean": 24.22,
        "std": 2.18,
        "min": 22.35,
        "max": 28.25,
        "runs": [
          28.25,
          24.75,
          23.2,
          22.55,
          22.35
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Physics_003": {
        "score": 53.7,
        "mean": 46.18,
        "std": 6.11,
        "min": 39.1,
        "max": 53.7,
        "runs": [
          53.7,
          52.6,
          45.5,
          40.0,
          39.1
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GPT-5.5",
        "model_display": "GPT-5.5",
        "details_exported": false,
        "pass_mode": "pass5"
      }
    },
    "ResearchHarness (MiMo-V2.5)": {
      "Astronomy_000": {
        "score": 20.0,
        "mean": 13.74,
        "std": 4.59,
        "min": 9.5,
        "max": 20.0,
        "runs": [
          20.0,
          18.6,
          11.0,
          9.6,
          9.5
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Astronomy_001": {
        "score": 38.4,
        "mean": 22.68,
        "std": 16.1,
        "min": 0.0,
        "max": 38.4,
        "runs": [
          38.4,
          38.0,
          30.0,
          7.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Astronomy_002": {
        "score": 26.0,
        "mean": 22.58,
        "std": 2.21,
        "min": 20.0,
        "max": 26.0,
        "runs": [
          26.0,
          24.1,
          22.1,
          20.7,
          20.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Astronomy_003": {
        "score": 48.9,
        "mean": 47.78,
        "std": 0.82,
        "min": 46.4,
        "max": 48.9,
        "runs": [
          48.9,
          48.2,
          47.8,
          47.6,
          46.4
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Chemistry_000": {
        "score": 19.7,
        "mean": 14.54,
        "std": 3.49,
        "min": 10.45,
        "max": 19.7,
        "runs": [
          19.7,
          17.55,
          12.7,
          12.3,
          10.45
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Chemistry_001": {
        "score": 13.0,
        "mean": 7.16,
        "std": 3.63,
        "min": 2.75,
        "max": 13.0,
        "runs": [
          13.0,
          9.05,
          6.75,
          4.25,
          2.75
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Chemistry_002": {
        "score": 0.0,
        "mean": 0.0,
        "std": 0.0,
        "min": 0.0,
        "max": 0.0,
        "runs": [
          0.0,
          0.0,
          0.0,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Chemistry_003": {
        "score": 13.6,
        "mean": 8.74,
        "std": 3.65,
        "min": 3.0,
        "max": 13.6,
        "runs": [
          13.6,
          10.7,
          9.9,
          6.5,
          3.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Earth_000": {
        "score": 21.5,
        "mean": 18.66,
        "std": 2.15,
        "min": 15.4,
        "max": 21.5,
        "runs": [
          21.5,
          19.8,
          19.5,
          17.1,
          15.4
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Earth_001": {
        "score": 40.45,
        "mean": 34.0,
        "std": 6.45,
        "min": 22.44,
        "max": 40.45,
        "runs": [
          40.45,
          39.4,
          35.18,
          32.55,
          22.44
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Earth_002": {
        "score": 24.9,
        "mean": 21.64,
        "std": 3.65,
        "min": 14.6,
        "max": 24.9,
        "runs": [
          24.9,
          23.9,
          22.4,
          22.4,
          14.6
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Earth_003": {
        "score": 5.1,
        "mean": 2.88,
        "std": 2.05,
        "min": 0.0,
        "max": 5.1,
        "runs": [
          5.1,
          4.9,
          3.4,
          1.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Energy_000": {
        "score": 19.0,
        "mean": 11.38,
        "std": 5.83,
        "min": 3.0,
        "max": 19.0,
        "runs": [
          19.0,
          16.9,
          9.0,
          9.0,
          3.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Energy_001": {
        "score": 18.2,
        "mean": 11.4,
        "std": 6.07,
        "min": 0.0,
        "max": 18.2,
        "runs": [
          18.2,
          13.6,
          12.7,
          12.5,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Energy_002": {
        "score": 39.7,
        "mean": 33.2,
        "std": 4.49,
        "min": 26.8,
        "max": 39.7,
        "runs": [
          39.7,
          35.8,
          33.75,
          29.95,
          26.8
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Energy_003": {
        "score": 27.3,
        "mean": 19.42,
        "std": 7.43,
        "min": 6.0,
        "max": 27.3,
        "runs": [
          27.3,
          24.3,
          22.0,
          17.5,
          6.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Information_000": {
        "score": 31.5,
        "mean": 16.38,
        "std": 9.82,
        "min": 2.0,
        "max": 31.5,
        "runs": [
          31.5,
          21.4,
          15.0,
          12.0,
          2.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Information_001": {
        "score": 6.4,
        "mean": 4.0,
        "std": 1.64,
        "min": 1.6,
        "max": 6.4,
        "runs": [
          6.4,
          5.0,
          4.0,
          3.0,
          1.6
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Information_002": {
        "score": 35.3,
        "mean": 16.56,
        "std": 10.71,
        "min": 6.2,
        "max": 35.3,
        "runs": [
          35.3,
          19.2,
          15.9,
          6.2,
          6.2
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Information_003": {
        "score": 14.35,
        "mean": 10.19,
        "std": 2.45,
        "min": 7.4,
        "max": 14.35,
        "runs": [
          14.35,
          11.2,
          9.7,
          8.3,
          7.4
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Life_000": {
        "score": 8.5,
        "mean": 6.29,
        "std": 1.36,
        "min": 4.8,
        "max": 8.5,
        "runs": [
          8.5,
          6.7,
          6.55,
          4.9,
          4.8
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Life_001": {
        "score": 13.65,
        "mean": 10.64,
        "std": 2.18,
        "min": 7.2,
        "max": 13.65,
        "runs": [
          13.65,
          12.15,
          10.25,
          9.95,
          7.2
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Life_002": {
        "score": 10.1,
        "mean": 5.52,
        "std": 4.6,
        "min": 0.0,
        "max": 10.1,
        "runs": [
          10.1,
          10.0,
          7.5,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Life_003": {
        "score": 32.0,
        "mean": 24.34,
        "std": 4.75,
        "min": 18.0,
        "max": 32.0,
        "runs": [
          32.0,
          25.4,
          25.4,
          20.9,
          18.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Material_000": {
        "score": 20.5,
        "mean": 12.58,
        "std": 6.3,
        "min": 3.6,
        "max": 20.5,
        "runs": [
          20.5,
          16.9,
          14.9,
          7.0,
          3.6
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Material_001": {
        "score": 13.75,
        "mean": 10.44,
        "std": 2.57,
        "min": 7.0,
        "max": 13.75,
        "runs": [
          13.75,
          12.0,
          11.55,
          7.9,
          7.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Material_002": {
        "score": 35.44,
        "mean": 23.34,
        "std": 6.68,
        "min": 14.85,
        "max": 35.44,
        "runs": [
          35.44,
          22.72,
          21.99,
          21.7,
          14.85
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Material_003": {
        "score": 18.05,
        "mean": 15.77,
        "std": 2.23,
        "min": 12.75,
        "max": 18.05,
        "runs": [
          18.05,
          17.9,
          16.65,
          13.5,
          12.75
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Math_000": {
        "score": 27.05,
        "mean": 15.49,
        "std": 6.85,
        "min": 7.95,
        "max": 27.05,
        "runs": [
          27.05,
          19.2,
          12.25,
          11.0,
          7.95
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Math_001": {
        "score": 32.2,
        "mean": 24.6,
        "std": 4.97,
        "min": 19.6,
        "max": 32.2,
        "runs": [
          32.2,
          28.8,
          21.4,
          21.0,
          19.6
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Math_002": {
        "score": 18.6,
        "mean": 12.58,
        "std": 5.24,
        "min": 3.9,
        "max": 18.6,
        "runs": [
          18.6,
          15.5,
          15.4,
          9.5,
          3.9
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Math_003": {
        "score": 36.75,
        "mean": 14.55,
        "std": 11.21,
        "min": 6.0,
        "max": 36.75,
        "runs": [
          36.75,
          10.0,
          10.0,
          10.0,
          6.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Neuroscience_000": {
        "score": 10.0,
        "mean": 5.44,
        "std": 2.96,
        "min": 1.6,
        "max": 10.0,
        "runs": [
          10.0,
          7.0,
          5.6,
          3.0,
          1.6
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Neuroscience_001": {
        "score": 7.5,
        "mean": 3.74,
        "std": 2.05,
        "min": 1.75,
        "max": 7.5,
        "runs": [
          7.5,
          4.25,
          2.7,
          2.5,
          1.75
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Neuroscience_002": {
        "score": 3.0,
        "mean": 1.56,
        "std": 0.95,
        "min": 0.3,
        "max": 3.0,
        "runs": [
          3.0,
          2.0,
          1.75,
          0.75,
          0.3
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Neuroscience_003": {
        "score": 13.8,
        "mean": 5.82,
        "std": 4.46,
        "min": 2.25,
        "max": 13.8,
        "runs": [
          13.8,
          7.65,
          2.7,
          2.7,
          2.25
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Physics_000": {
        "score": 26.9,
        "mean": 23.38,
        "std": 3.04,
        "min": 19.0,
        "max": 26.9,
        "runs": [
          26.9,
          26.1,
          24.1,
          20.8,
          19.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Physics_001": {
        "score": 27.5,
        "mean": 24.37,
        "std": 2.41,
        "min": 21.05,
        "max": 27.5,
        "runs": [
          27.5,
          26.5,
          24.35,
          22.45,
          21.05
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Physics_002": {
        "score": 29.6,
        "mean": 26.86,
        "std": 2.27,
        "min": 23.7,
        "max": 29.6,
        "runs": [
          29.6,
          28.45,
          27.85,
          24.7,
          23.7
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Physics_003": {
        "score": 53.5,
        "mean": 43.62,
        "std": 5.74,
        "min": 37.5,
        "max": 53.5,
        "runs": [
          53.5,
          46.1,
          41.9,
          39.1,
          37.5
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "MiMo-V2.5",
        "model_display": "MiMo-V2.5",
        "details_exported": false,
        "pass_mode": "pass5"
      }
    },
    "ResearchHarness (GLM-5.1)": {
      "Astronomy_000": {
        "score": 29.4,
        "mean": 19.56,
        "std": 6.76,
        "min": 11.5,
        "max": 29.4,
        "runs": [
          29.4,
          24.4,
          19.5,
          13.0,
          11.5
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Astronomy_001": {
        "score": 39.4,
        "mean": 29.4,
        "std": 6.22,
        "min": 20.4,
        "max": 39.4,
        "runs": [
          39.4,
          31.2,
          29.6,
          26.4,
          20.4
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Astronomy_002": {
        "score": 28.9,
        "mean": 21.1,
        "std": 10.69,
        "min": 0.0,
        "max": 28.9,
        "runs": [
          28.9,
          27.0,
          26.0,
          23.6,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Astronomy_003": {
        "score": 46.3,
        "mean": 17.94,
        "std": 21.99,
        "min": 0.0,
        "max": 46.3,
        "runs": [
          46.3,
          43.4,
          0.0,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Chemistry_000": {
        "score": 22.65,
        "mean": 6.86,
        "std": 9.09,
        "min": 0.0,
        "max": 22.65,
        "runs": [
          22.65,
          11.65,
          0.0,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Chemistry_001": {
        "score": 0.0,
        "mean": 0.0,
        "std": 0.0,
        "min": 0.0,
        "max": 0.0,
        "runs": [
          0.0,
          0.0,
          0.0,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Chemistry_002": {
        "score": 0.3,
        "mean": 0.06,
        "std": 0.12,
        "min": 0.0,
        "max": 0.3,
        "runs": [
          0.3,
          0.0,
          0.0,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Chemistry_003": {
        "score": 24.6,
        "mean": 12.48,
        "std": 10.65,
        "min": 0.0,
        "max": 24.6,
        "runs": [
          24.6,
          22.5,
          15.3,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Earth_000": {
        "score": 24.8,
        "mean": 4.96,
        "std": 9.92,
        "min": 0.0,
        "max": 24.8,
        "runs": [
          24.8,
          0.0,
          0.0,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Earth_001": {
        "score": 0.0,
        "mean": 0.0,
        "std": 0.0,
        "min": 0.0,
        "max": 0.0,
        "runs": [
          0.0,
          0.0,
          0.0,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Earth_002": {
        "score": 0.0,
        "mean": 0.0,
        "std": 0.0,
        "min": 0.0,
        "max": 0.0,
        "runs": [
          0.0,
          0.0,
          0.0,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Earth_003": {
        "score": 0.0,
        "mean": 0.0,
        "std": 0.0,
        "min": 0.0,
        "max": 0.0,
        "runs": [
          0.0,
          0.0,
          0.0,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Energy_000": {
        "score": 20.8,
        "mean": 9.44,
        "std": 8.52,
        "min": 0.0,
        "max": 20.8,
        "runs": [
          20.8,
          16.9,
          9.5,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Energy_001": {
        "score": 0.0,
        "mean": 0.0,
        "std": 0.0,
        "min": 0.0,
        "max": 0.0,
        "runs": [
          0.0,
          0.0,
          0.0,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Energy_002": {
        "score": 0.0,
        "mean": 0.0,
        "std": 0.0,
        "min": 0.0,
        "max": 0.0,
        "runs": [
          0.0,
          0.0,
          0.0,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Energy_003": {
        "score": 0.0,
        "mean": 0.0,
        "std": 0.0,
        "min": 0.0,
        "max": 0.0,
        "runs": [
          0.0,
          0.0,
          0.0,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Information_000": {
        "score": 0.0,
        "mean": 0.0,
        "std": 0.0,
        "min": 0.0,
        "max": 0.0,
        "runs": [
          0.0,
          0.0,
          0.0,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Information_001": {
        "score": 0.0,
        "mean": 0.0,
        "std": 0.0,
        "min": 0.0,
        "max": 0.0,
        "runs": [
          0.0,
          0.0,
          0.0,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Information_002": {
        "score": 35.6,
        "mean": 18.06,
        "std": 11.62,
        "min": 5.6,
        "max": 35.6,
        "runs": [
          35.6,
          27.8,
          12.6,
          8.7,
          5.6
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Information_003": {
        "score": 6.85,
        "mean": 2.63,
        "std": 3.23,
        "min": 0.0,
        "max": 6.85,
        "runs": [
          6.85,
          6.3,
          0.0,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Life_000": {
        "score": 11.25,
        "mean": 9.01,
        "std": 1.9,
        "min": 6.05,
        "max": 11.25,
        "runs": [
          11.25,
          10.85,
          8.8,
          8.1,
          6.05
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Life_001": {
        "score": 23.3,
        "mean": 9.89,
        "std": 7.61,
        "min": 0.0,
        "max": 23.3,
        "runs": [
          23.3,
          10.95,
          7.75,
          7.45,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Life_002": {
        "score": 9.6,
        "mean": 5.78,
        "std": 3.12,
        "min": 0.6,
        "max": 9.6,
        "runs": [
          9.6,
          8.3,
          5.4,
          5.0,
          0.6
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Life_003": {
        "score": 40.8,
        "mean": 31.92,
        "std": 9.22,
        "min": 14.3,
        "max": 40.8,
        "runs": [
          40.8,
          36.2,
          36.1,
          32.2,
          14.3
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Material_000": {
        "score": 19.3,
        "mean": 15.94,
        "std": 3.49,
        "min": 9.9,
        "max": 19.3,
        "runs": [
          19.3,
          19.0,
          17.1,
          14.4,
          9.9
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Material_001": {
        "score": 25.9,
        "mean": 15.53,
        "std": 6.5,
        "min": 9.5,
        "max": 25.9,
        "runs": [
          25.9,
          20.45,
          11.3,
          10.5,
          9.5
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Material_002": {
        "score": 37.11,
        "mean": 20.93,
        "std": 12.12,
        "min": 0.0,
        "max": 37.11,
        "runs": [
          37.11,
          25.54,
          23.55,
          18.45,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Material_003": {
        "score": 19.2,
        "mean": 15.82,
        "std": 2.53,
        "min": 11.85,
        "max": 19.2,
        "runs": [
          19.2,
          17.15,
          16.65,
          14.25,
          11.85
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Math_000": {
        "score": 32.45,
        "mean": 27.53,
        "std": 3.52,
        "min": 23.55,
        "max": 32.45,
        "runs": [
          32.45,
          30.6,
          27.0,
          24.05,
          23.55
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Math_001": {
        "score": 43.9,
        "mean": 26.88,
        "std": 8.84,
        "min": 18.9,
        "max": 43.9,
        "runs": [
          43.9,
          26.3,
          23.4,
          21.9,
          18.9
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Math_002": {
        "score": 10.9,
        "mean": 9.22,
        "std": 1.41,
        "min": 6.8,
        "max": 10.9,
        "runs": [
          10.9,
          10.2,
          9.5,
          8.7,
          6.8
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Math_003": {
        "score": 19.75,
        "mean": 11.31,
        "std": 6.64,
        "min": 0.0,
        "max": 19.75,
        "runs": [
          19.75,
          15.75,
          11.05,
          10.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Neuroscience_000": {
        "score": 17.0,
        "mean": 10.32,
        "std": 4.6,
        "min": 5.0,
        "max": 17.0,
        "runs": [
          17.0,
          14.0,
          9.6,
          6.0,
          5.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Neuroscience_001": {
        "score": 9.2,
        "mean": 4.39,
        "std": 3.04,
        "min": 0.0,
        "max": 9.2,
        "runs": [
          9.2,
          5.5,
          4.5,
          2.75,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Neuroscience_002": {
        "score": 4.5,
        "mean": 1.75,
        "std": 1.96,
        "min": 0.0,
        "max": 4.5,
        "runs": [
          4.5,
          3.75,
          0.5,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Neuroscience_003": {
        "score": 15.8,
        "mean": 8.15,
        "std": 5.37,
        "min": 0.0,
        "max": 15.8,
        "runs": [
          15.8,
          11.85,
          6.95,
          6.15,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Physics_000": {
        "score": 21.8,
        "mean": 4.36,
        "std": 8.72,
        "min": 0.0,
        "max": 21.8,
        "runs": [
          21.8,
          0.0,
          0.0,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Physics_001": {
        "score": 0.0,
        "mean": 0.0,
        "std": 0.0,
        "min": 0.0,
        "max": 0.0,
        "runs": [
          0.0,
          0.0,
          0.0,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Physics_002": {
        "score": 0.0,
        "mean": 0.0,
        "std": 0.0,
        "min": 0.0,
        "max": 0.0,
        "runs": [
          0.0,
          0.0,
          0.0,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      },
      "Physics_003": {
        "score": 42.0,
        "mean": 16.6,
        "std": 20.33,
        "min": 0.0,
        "max": 42.0,
        "runs": [
          42.0,
          41.0,
          0.0,
          0.0,
          0.0
        ],
        "attempts": 5,
        "scored_runs": 5,
        "model": "GLM-5.1",
        "model_display": "GLM-5.1",
        "details_exported": false,
        "pass_mode": "pass5"
      }
    }
  },
  "frontier": {
    "Astronomy_000": 32.6,
    "Astronomy_001": 40.6,
    "Astronomy_002": 29.9,
    "Astronomy_003": 50.8,
    "Chemistry_000": 34.0,
    "Chemistry_001": 13.0,
    "Chemistry_002": 18.2,
    "Chemistry_003": 30.6,
    "Earth_000": 24.8,
    "Earth_001": 44.23,
    "Earth_002": 36.1,
    "Earth_003": 13.0,
    "Energy_000": 25.0,
    "Energy_001": 31.7,
    "Energy_002": 40.9,
    "Energy_003": 44.0,
    "Information_000": 58.6,
    "Information_001": 19.6,
    "Information_002": 45.3,
    "Information_003": 21.85,
    "Life_000": 12.55,
    "Life_001": 29.0,
    "Life_002": 24.0,
    "Life_003": 56.5,
    "Material_000": 36.8,
    "Material_001": 31.05,
    "Material_002": 43.71,
    "Material_003": 28.8,
    "Math_000": 45.0,
    "Math_001": 46.4,
    "Math_002": 26.5,
    "Math_003": 36.75,
    "Neuroscience_000": 19.0,
    "Neuroscience_001": 25.95,
    "Neuroscience_002": 6.25,
    "Neuroscience_003": 20.75,
    "Physics_000": 37.8,
    "Physics_001": 37.65,
    "Physics_002": 41.2,
    "Physics_003": 64.4
  }
}
