{
  "source": "CellScientist manuscript, Table 3",
  "metric": "Held-out global Pearson correlation (global PCC; higher is better)",
  "seed_count": 5,
  "dispersion": "Mean \u00b1 sample standard deviation over five paired controller and fitting seeds within each task/split setting.",
  "budget_definition": "Each B counts evaluated candidates including the shared initial candidate h0; B=3 and B=5 therefore allow two and four revisions.",
  "rows": [
    {
      "task_id": "BBBC036_plate",
      "dataset": "BBBC036",
      "split": "Plate",
      "budget": 3,
      "results": {
        "fixed": {
          "mean": 0.1664,
          "std": 0.0005,
          "display": "0.1664 \u00b1 0.0005"
        },
        "flat": {
          "mean": 0.1828,
          "std": 0.0005,
          "display": "0.1828 \u00b1 0.0005"
        },
        "random": {
          "mean": 0.225,
          "std": 0.0493,
          "display": "0.2250 \u00b1 0.0493"
        },
        "aide": {
          "mean": 0.2244,
          "std": 0.0276,
          "display": "0.2244 \u00b1 0.0276"
        },
        "full": {
          "mean": 0.265,
          "std": 0.0005,
          "display": "0.2650 \u00b1 0.0005"
        }
      }
    },
    {
      "task_id": "BBBC036_plate",
      "dataset": "BBBC036",
      "split": "Plate",
      "budget": 5,
      "results": {
        "fixed": {
          "mean": 0.1664,
          "std": 0.0005,
          "display": "0.1664 \u00b1 0.0005"
        },
        "flat": {
          "mean": 0.2073,
          "std": 0.0056,
          "display": "0.2073 \u00b1 0.0056"
        },
        "random": {
          "mean": 0.2277,
          "std": 0.0502,
          "display": "0.2277 \u00b1 0.0502"
        },
        "aide": {
          "mean": 0.2477,
          "std": 0.0294,
          "display": "0.2477 \u00b1 0.0294"
        },
        "full": {
          "mean": 0.2674,
          "std": 0.0003,
          "display": "0.2674 \u00b1 0.0003"
        }
      }
    },
    {
      "task_id": "BBBC036_plate",
      "dataset": "BBBC036",
      "split": "Plate",
      "budget": 10,
      "results": {
        "fixed": {
          "mean": 0.1664,
          "std": 0.0005,
          "display": "0.1664 \u00b1 0.0005"
        },
        "flat": {
          "mean": 0.2175,
          "std": 0.0061,
          "display": "0.2175 \u00b1 0.0061"
        },
        "random": {
          "mean": 0.2626,
          "std": 0.0077,
          "display": "0.2626 \u00b1 0.0077"
        },
        "aide": {
          "mean": 0.2562,
          "std": 0.0205,
          "display": "0.2562 \u00b1 0.0205"
        },
        "full": {
          "mean": 0.2693,
          "std": 0.0012,
          "display": "0.2693 \u00b1 0.0012"
        }
      }
    },
    {
      "task_id": "BBBC036_smiles",
      "dataset": "BBBC036",
      "split": "SMILES",
      "budget": 3,
      "results": {
        "fixed": {
          "mean": 0.1547,
          "std": 0.0005,
          "display": "0.1547 \u00b1 0.0005"
        },
        "flat": {
          "mean": 0.1583,
          "std": 0.0046,
          "display": "0.1583 \u00b1 0.0046"
        },
        "random": {
          "mean": 0.2052,
          "std": 0.0413,
          "display": "0.2052 \u00b1 0.0413"
        },
        "aide": {
          "mean": 0.2058,
          "std": 0.023,
          "display": "0.2058 \u00b1 0.0230"
        },
        "full": {
          "mean": 0.2336,
          "std": 0.0016,
          "display": "0.2336 \u00b1 0.0016"
        }
      }
    },
    {
      "task_id": "BBBC036_smiles",
      "dataset": "BBBC036",
      "split": "SMILES",
      "budget": 5,
      "results": {
        "fixed": {
          "mean": 0.1547,
          "std": 0.0005,
          "display": "0.1547 \u00b1 0.0005"
        },
        "flat": {
          "mean": 0.1882,
          "std": 0.0129,
          "display": "0.1882 \u00b1 0.0129"
        },
        "random": {
          "mean": 0.2049,
          "std": 0.0417,
          "display": "0.2049 \u00b1 0.0417"
        },
        "aide": {
          "mean": 0.2231,
          "std": 0.0169,
          "display": "0.2231 \u00b1 0.0169"
        },
        "full": {
          "mean": 0.2346,
          "std": 0.0009,
          "display": "0.2346 \u00b1 0.0009"
        }
      }
    },
    {
      "task_id": "BBBC036_smiles",
      "dataset": "BBBC036",
      "split": "SMILES",
      "budget": 10,
      "results": {
        "fixed": {
          "mean": 0.1547,
          "std": 0.0005,
          "display": "0.1547 \u00b1 0.0005"
        },
        "flat": {
          "mean": 0.1948,
          "std": 0.0005,
          "display": "0.1948 \u00b1 0.0005"
        },
        "random": {
          "mean": 0.2347,
          "std": 0.0016,
          "display": "0.2347 \u00b1 0.0016"
        },
        "aide": {
          "mean": 0.2235,
          "std": 0.0164,
          "display": "0.2235 \u00b1 0.0164"
        },
        "full": {
          "mean": 0.2346,
          "std": 0.0009,
          "display": "0.2346 \u00b1 0.0009"
        }
      }
    },
    {
      "task_id": "BBBC047_plate",
      "dataset": "BBBC047",
      "split": "Plate",
      "budget": 3,
      "results": {
        "fixed": {
          "mean": 0.2442,
          "std": 0.0008,
          "display": "0.2442 \u00b1 0.0008"
        },
        "flat": {
          "mean": 0.25,
          "std": 0.0055,
          "display": "0.2500 \u00b1 0.0055"
        },
        "random": {
          "mean": 0.3032,
          "std": 0.05,
          "display": "0.3032 \u00b1 0.0500"
        },
        "aide": {
          "mean": 0.3013,
          "std": 0.0381,
          "display": "0.3013 \u00b1 0.0381"
        },
        "full": {
          "mean": 0.3396,
          "std": 0.0004,
          "display": "0.3396 \u00b1 0.0004"
        }
      }
    },
    {
      "task_id": "BBBC047_plate",
      "dataset": "BBBC047",
      "split": "Plate",
      "budget": 5,
      "results": {
        "fixed": {
          "mean": 0.2442,
          "std": 0.0008,
          "display": "0.2442 \u00b1 0.0008"
        },
        "flat": {
          "mean": 0.263,
          "std": 0.013,
          "display": "0.2630 \u00b1 0.0130"
        },
        "random": {
          "mean": 0.3066,
          "std": 0.0514,
          "display": "0.3066 \u00b1 0.0514"
        },
        "aide": {
          "mean": 0.3153,
          "std": 0.0399,
          "display": "0.3153 \u00b1 0.0399"
        },
        "full": {
          "mean": 0.3457,
          "std": 0.0004,
          "display": "0.3457 \u00b1 0.0004"
        }
      }
    },
    {
      "task_id": "BBBC047_plate",
      "dataset": "BBBC047",
      "split": "Plate",
      "budget": 10,
      "results": {
        "fixed": {
          "mean": 0.2442,
          "std": 0.0008,
          "display": "0.2442 \u00b1 0.0008"
        },
        "flat": {
          "mean": 0.2777,
          "std": 0.0005,
          "display": "0.2777 \u00b1 0.0005"
        },
        "random": {
          "mean": 0.3458,
          "std": 0.0027,
          "display": "0.3458 \u00b1 0.0027"
        },
        "aide": {
          "mean": 0.3191,
          "std": 0.0372,
          "display": "0.3191 \u00b1 0.0372"
        },
        "full": {
          "mean": 0.3481,
          "std": 0.0009,
          "display": "0.3481 \u00b1 0.0009"
        }
      }
    },
    {
      "task_id": "BBBC047_smiles",
      "dataset": "BBBC047",
      "split": "SMILES",
      "budget": 3,
      "results": {
        "fixed": {
          "mean": 0.2504,
          "std": 0.0004,
          "display": "0.2504 \u00b1 0.0004"
        },
        "flat": {
          "mean": 0.2651,
          "std": 0.0002,
          "display": "0.2651 \u00b1 0.0002"
        },
        "random": {
          "mean": 0.3131,
          "std": 0.0536,
          "display": "0.3131 \u00b1 0.0536"
        },
        "aide": {
          "mean": 0.3123,
          "std": 0.0346,
          "display": "0.3123 \u00b1 0.0346"
        },
        "full": {
          "mean": 0.3529,
          "std": 0.001,
          "display": "0.3529 \u00b1 0.0010"
        }
      }
    },
    {
      "task_id": "BBBC047_smiles",
      "dataset": "BBBC047",
      "split": "SMILES",
      "budget": 5,
      "results": {
        "fixed": {
          "mean": 0.2504,
          "std": 0.0004,
          "display": "0.2504 \u00b1 0.0004"
        },
        "flat": {
          "mean": 0.2916,
          "std": 0.0003,
          "display": "0.2916 \u00b1 0.0003"
        },
        "random": {
          "mean": 0.3207,
          "std": 0.0582,
          "display": "0.3207 \u00b1 0.0582"
        },
        "aide": {
          "mean": 0.3371,
          "std": 0.0417,
          "display": "0.3371 \u00b1 0.0417"
        },
        "full": {
          "mean": 0.3614,
          "std": 0.0003,
          "display": "0.3614 \u00b1 0.0003"
        }
      }
    },
    {
      "task_id": "BBBC047_smiles",
      "dataset": "BBBC047",
      "split": "SMILES",
      "budget": 10,
      "results": {
        "fixed": {
          "mean": 0.2504,
          "std": 0.0004,
          "display": "0.2504 \u00b1 0.0004"
        },
        "flat": {
          "mean": 0.292,
          "std": 0.0003,
          "display": "0.2920 \u00b1 0.0003"
        },
        "random": {
          "mean": 0.3668,
          "std": 0.0073,
          "display": "0.3668 \u00b1 0.0073"
        },
        "aide": {
          "mean": 0.356,
          "std": 0.0359,
          "display": "0.3560 \u00b1 0.0359"
        },
        "full": {
          "mean": 0.3751,
          "std": 0.0017,
          "display": "0.3751 \u00b1 0.0017"
        }
      }
    }
  ],
  "effects": {
    "metric": "Difference in held-out global PCC (CellScientist \u2212 Random routing)",
    "source": "Table 3 lower panel; Appendix C.4 statistical aggregation",
    "ci_level": 0.95,
    "ci_method": "Paired-setting t-confidence interval across four task/split settings. Effects are first averaged over the five paired seeds within each setting; intervals are computed from unrounded run-level values.",
    "rows": [
      {
        "budget": 3,
        "mean": 0.0361,
        "lower": 0.0275,
        "upper": 0.0448,
        "display": "+0.0361 [0.0275, 0.0448]"
      },
      {
        "budget": 5,
        "mean": 0.0373,
        "lower": 0.0292,
        "upper": 0.0454,
        "display": "+0.0373 [0.0292, 0.0454]"
      },
      {
        "budget": 10,
        "mean": 0.0043,
        "lower": -0.0018,
        "upper": 0.0104,
        "display": "+0.0043 [\u22120.0018, 0.0104]"
      }
    ],
    "headline": "Higher held-out PCC after three and five candidate evaluations.",
    "interpretation": "The strongest supported result is improved candidate-evaluation efficiency at small budgets. At ten evaluations the mean gap narrows and its 95% CI includes zero.",
    "chart_caption": "CellScientist minus Random routing, averaged across four task/split settings. Bars show paired-setting 95% t-confidence intervals."
  }
}