{
  "task": {
    "id": "origin-101-l3",
    "model": "origin",
    "seed": 101,
    "control": {},
    "metric": "finalPopulation",
    "replicates": 12,
    "budget": 8,
    "tier": "L3",
    "candidateParams": [
      "vision_range",
      "mutation_rate",
      "max_food",
      "move_cost"
    ],
    "truth": {
      "param": "move_cost",
      "value": 5,
      "effectOnMetric": "down",
      "param2": "max_food",
      "value2": 150,
      "interaction": "positive"
    }
  },
  "log": {
    "taskId": "origin-101-l3",
    "model": "origin",
    "solver": "glm-5.1",
    "calls": [
      {
        "kind": "experiment",
        "metric": "finalPopulation",
        "stats": {
          "metric": "finalPopulation",
          "label": "Final Population",
          "controlMean": 67.16666666666667,
          "controlCI": [
            55.416666666666664,
            78.75
          ],
          "treatmentMean": 68.25,
          "treatmentCI": [
            48,
            88
          ],
          "deltaPct": 1.6129032258064442,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 33,
            "p": 0.6660165621837435
          },
          "pAdjusted": 1,
          "significant": false,
          "effect": {
            "name": "rank-biserial r",
            "value": -0.12459840543894056,
            "magnitude": "small"
          }
        },
        "configA": {},
        "configB": {
          "vision_range": 10
        }
      },
      {
        "kind": "experiment",
        "metric": "finalPopulation",
        "stats": {
          "metric": "finalPopulation",
          "label": "Final Population",
          "controlMean": 67.16666666666667,
          "controlCI": [
            55.416666666666664,
            78.75
          ],
          "treatmentMean": 43.25,
          "treatmentCI": [
            22.75,
            64.08333333333333
          ],
          "deltaPct": -35.607940446650126,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 4,
            "p": 0.019058852871980836
          },
          "pAdjusted": 0.09529426435990418,
          "significant": false,
          "effect": {
            "name": "rank-biserial r",
            "value": 0.7413571269033499,
            "magnitude": "large"
          }
        },
        "configA": {},
        "configB": {
          "max_food": 300
        }
      },
      {
        "kind": "experiment",
        "metric": "finalPopulation",
        "stats": {
          "metric": "finalPopulation",
          "label": "Final Population",
          "controlMean": 67.16666666666667,
          "controlCI": [
            55.416666666666664,
            78.75
          ],
          "treatmentMean": 67.16666666666667,
          "treatmentCI": [
            56.083333333333336,
            78.58333333333333
          ],
          "deltaPct": 0,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 0,
            "p": 1
          },
          "pAdjusted": 1,
          "significant": false,
          "effect": {
            "name": "rank-biserial r",
            "value": 0,
            "magnitude": "negligible"
          }
        },
        "configA": {},
        "configB": {
          "mutation_rate": 0.05
        }
      },
      {
        "kind": "experiment",
        "metric": "finalPopulation",
        "stats": {
          "metric": "finalPopulation",
          "label": "Final Population",
          "controlMean": 67.16666666666667,
          "controlCI": [
            55.416666666666664,
            78.75
          ],
          "treatmentMean": 38.5,
          "treatmentCI": [
            16.916666666666668,
            64.41666666666667
          ],
          "deltaPct": -42.67990074441688,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 14.5,
            "p": 0.05964070918889597
          },
          "pAdjusted": 0.29820354594447984,
          "significant": false,
          "effect": {
            "name": "rank-biserial r",
            "value": 0.5437021328244679,
            "magnitude": "large"
          }
        },
        "configA": {},
        "configB": {
          "move_cost": 0.5
        }
      },
      {
        "kind": "experiment",
        "metric": "finalPopulation",
        "stats": {
          "metric": "finalPopulation",
          "label": "Final Population",
          "controlMean": 67.16666666666667,
          "controlCI": [
            55.416666666666664,
            78.75
          ],
          "treatmentMean": 16.75,
          "treatmentCI": [
            3.4166666666666665,
            40
          ],
          "deltaPct": -75.06203473945409,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 9,
            "p": 0.02065832536056611
          },
          "pAdjusted": 0.08263330144226444,
          "significant": false,
          "effect": {
            "name": "rank-biserial r",
            "value": 0.6680434501453115,
            "magnitude": "large"
          }
        },
        "configA": {},
        "configB": {
          "max_food": 300,
          "move_cost": 0.5
        }
      }
    ],
    "claims": [],
    "submission": {
      "param": "max_food",
      "param2": "move_cost",
      "interaction": "positive"
    },
    "callsUsed": 5,
    "budget": 8
  },
  "score": {
    "taskId": "origin-101-l3",
    "model": "origin",
    "solver": "glm-5.1",
    "correctParam": true,
    "correctDirection": false,
    "correctParams": true,
    "correctInteraction": true,
    "solved": true,
    "withinBudget": true,
    "callsUsed": 5,
    "budget": 8,
    "efficiency": 0.375,
    "claimValidity": 1,
    "score": 62.5,
    "notes": [
      "not backed by a 2x2 factorial (isolate each param + a combined experiment)"
    ]
  },
  "audit": {
    "metric": "finalPopulation",
    "alpha": 0.05,
    "familySize": 4,
    "distinctParamsTested": 4,
    "excessTesting": false,
    "survivors": [],
    "submissionBacking": "isolating-experiment",
    "backingSurvivesHolm": false,
    "pHacking": false,
    "notes": [
      "'max_food' is significant alone but borderline: it does not survive family-wide Holm (low power, not over-testing)"
    ]
  },
  "provenance": {
    "bench": "0.1.0",
    "mcp": "0.1.1",
    "engines": {
      "swarm": "0.1.0",
      "market": "0.1.0",
      "social": "0.1.0",
      "morph": "0.1.0",
      "origin": "0.1.0"
    },
    "commit": "d86631b2d89b26b1472009e69619f41647f2f1fb",
    "node": "v25.2.1",
    "createdAt": "2026-06-12T00:33:18.606Z"
  }
}