{
  "n_tasks": 100,
  "fixed": {
    "no_think": {
      "accuracy": 0.76,
      "mean_think": 0.0,
      "commit_dist": {
        "no_think": 100
      },
      "false_visible_commit": 0.08
    },
    "think_256": {
      "accuracy": 0.87,
      "mean_think": 246.19,
      "commit_dist": {
        "think_256": 100
      },
      "false_visible_commit": 0.07
    },
    "think_512": {
      "accuracy": 0.87,
      "mean_think": 403.95,
      "commit_dist": {
        "think_512": 100
      },
      "false_visible_commit": 0.07
    },
    "think_1024": {
      "accuracy": 0.91,
      "mean_think": 507.49,
      "commit_dist": {
        "think_1024": 100
      },
      "false_visible_commit": 0.06
    },
    "think_2048": {
      "accuracy": 0.86,
      "mean_think": 595.55,
      "commit_dist": {
        "think_2048": 100
      },
      "false_visible_commit": 0.05
    },
    "think_unbudgeted": {
      "accuracy": 0.84,
      "mean_think": 472.67,
      "commit_dist": {
        "think_unbudgeted": 100
      },
      "false_visible_commit": 0.05
    }
  },
  "controllers": {
    "esc[256-512-1024]": {
      "accuracy": 0.89,
      "mean_think": 317.16,
      "commit_dist": {
        "think_256": 94,
        "think_512": 1,
        "think_1024": 5
      },
      "false_visible_commit": 0.08
    },
    "esc[256-512-1024] +continue": {
      "accuracy": 0.89,
      "mean_think": 276.97,
      "commit_dist": {
        "think_256": 94,
        "think_512": 1,
        "think_1024": 5
      },
      "false_visible_commit": 0.08
    },
    "esc[256-512-1024-2048]": {
      "accuracy": 0.89,
      "mean_think": 359.07,
      "commit_dist": {
        "think_256": 94,
        "think_512": 1,
        "think_1024": 2,
        "think_2048": 3
      },
      "false_visible_commit": 0.08
    },
    "esc[256-512-1024-2048] +continue": {
      "accuracy": 0.89,
      "mean_think": 298.34,
      "commit_dist": {
        "think_256": 94,
        "think_512": 1,
        "think_1024": 2,
        "think_2048": 3
      },
      "false_visible_commit": 0.08
    },
    "esc[nothink-256-512-1024]": {
      "accuracy": 0.87,
      "mean_think": 81.21,
      "commit_dist": {
        "no_think": 84,
        "think_256": 12,
        "think_512": 1,
        "think_1024": 3
      },
      "false_visible_commit": 0.11
    },
    "esc[nothink-256-512-1024] +continue": {
      "accuracy": 0.87,
      "mean_think": 56.38,
      "commit_dist": {
        "no_think": 84,
        "think_256": 12,
        "think_512": 1,
        "think_1024": 3
      },
      "false_visible_commit": 0.11
    },
    "esc[nothink-1024] (2-tier)": {
      "accuracy": 0.88,
      "mean_think": 113.22,
      "commit_dist": {
        "no_think": 84,
        "think_1024": 16
      },
      "false_visible_commit": 0.1
    },
    "esc[nothink-1024] (2-tier) +continue": {
      "accuracy": 0.88,
      "mean_think": 113.22,
      "commit_dist": {
        "no_think": 84,
        "think_1024": 16
      },
      "false_visible_commit": 0.1
    }
  },
  "oracle_ceiling": {
    "accuracy": 0.93,
    "mean_think": 131.91
  }
}