{
  "protocol": {
    "seed": 20261003,
    "repeats": 1000,
    "p": 0.6,
    "dev_tasks": 60,
    "test_tasks": 500,
    "candidate_sizes": [
      1,
      8,
      32,
      128
    ],
    "selection": "dev argmax; first tie; candidates fixed before evaluation",
    "scope": "independent synthetic noise, no real ability gain, no RSI replication"
  },
  "environment": {
    "python": "3.14.6",
    "numpy": "2.5.3",
    "matplotlib": "3.11.2"
  },
  "results": [
    {
      "candidates": 1,
      "dev": {
        "mean": 0.5992000000000001,
        "mc_se": 0.0019493143648882869
      },
      "test": {
        "mean": 0.600308,
        "mc_se": 0.000709704739181596
      },
      "gap": {
        "mean": -0.0011079999999999786,
        "mc_se": 0.0020591141186491546
      },
      "gates": {
        "fixed_075": {
          "false_adoption_rate": 0.01,
          "mc_se": 0.003146426544510455,
          "wilson95": [
            0.005440754445529249,
            0.018309468870314774
          ],
          "theory": 0.010948230401696235
        },
        "uncorrected_005": {
          "false_adoption_rate": 0.038,
          "mc_se": 0.006046155803483731,
          "wilson95": [
            0.02780900450930263,
            0.05172692033135032
          ],
          "theory": 0.041288218812184585
        },
        "bonferroni": {
          "false_adoption_rate": 0.038,
          "mc_se": 0.006046155803483731,
          "wilson95": [
            0.02780900450930263,
            0.05172692033135032
          ],
          "theory": 0.041288218812184585
        }
      }
    },
    {
      "candidates": 8,
      "dev": {
        "mean": 0.68925,
        "mc_se": 0.0011517801138950512
      },
      "test": {
        "mean": 0.6010160000000001,
        "mc_se": 0.0006894832348441342
      },
      "gap": {
        "mean": 0.088234,
        "mc_se": 0.0013496271352009554
      },
      "gates": {
        "fixed_075": {
          "false_adoption_rate": 0.071,
          "mc_se": 0.008121514637061242,
          "wilson95": [
            0.056669695350451466,
            0.08861366343015481
          ],
          "theory": 0.08430214986447848
        },
        "uncorrected_005": {
          "false_adoption_rate": 0.306,
          "mc_se": 0.0145727142289966,
          "wilson95": [
            0.2782254333203639,
            0.3352593489720314
          ],
          "theory": 0.2863183768957758
        },
        "bonferroni": {
          "false_adoption_rate": 0.034,
          "mc_se": 0.005730968504537431,
          "wilson95": [
            0.024431348663226313,
            0.04713519024531107
          ],
          "theory": 0.03939343816584606
        }
      }
    },
    {
      "candidates": 32,
      "dev": {
        "mean": 0.7281499999999999,
        "mc_se": 0.0009158023650054241
      },
      "test": {
        "mean": 0.6000620000000001,
        "mc_se": 0.0006788143670731678
      },
      "gap": {
        "mean": 0.12808800000000004,
        "mc_se": 0.0011326474550222947
      },
      "gates": {
        "fixed_075": {
          "false_adoption_rate": 0.296,
          "mc_se": 0.01443551176785915,
          "wilson95": [
            0.26853097433797923,
            0.32503034312412726
          ],
          "theory": 0.29691346904378113
        },
        "uncorrected_005": {
          "false_adoption_rate": 0.75,
          "mc_se": 0.013693063937629153,
          "wilson95": [
            0.7222397197409139,
            0.7758469010163086
          ],
          "theory": 0.7405712779841441
        },
        "bonferroni": {
          "false_adoption_rate": 0.024,
          "mc_se": 0.004839834707921336,
          "wilson95": [
            0.016180168466636995,
            0.03546290561161149
          ],
          "theory": 0.02558746202654394
        }
      }
    },
    {
      "candidates": 128,
      "dev": {
        "mean": 0.7585,
        "mc_se": 0.0007295005551395
      },
      "test": {
        "mean": 0.6005199999999999,
        "mc_se": 0.0007060397241601158
      },
      "gap": {
        "mean": 0.15798,
        "mc_se": 0.00100040323312706
      },
      "gates": {
        "fixed_075": {
          "false_adoption_rate": 0.762,
          "mc_se": 0.01346684818359515,
          "wilson95": [
            0.7346343310291412,
            0.787360447524428
          ],
          "theory": 0.7556371886854376
        },
        "uncorrected_005": {
          "false_adoption_rate": 0.998,
          "mc_se": 0.0014127986409959495,
          "wilson95": [
            0.9927371921365079,
            0.9994513564118793
          ],
          "theory": 0.995470270950168
        },
        "bonferroni": {
          "false_adoption_rate": 0.022,
          "mc_se": 0.0046385342512479085,
          "wilson95": [
            0.014572707764791058,
            0.03308567334739963
          ],
          "theory": 0.03552744700240007
        }
      }
    }
  ]
}