{
  "task": {
    "id": "origin-202",
    "model": "origin",
    "seed": 202,
    "control": {},
    "metric": "finalPopulation",
    "replicates": 12,
    "budget": 8,
    "candidateParams": [
      "vision_range",
      "attack_damage",
      "reproduce_threshold"
    ],
    "truth": {
      "param": "reproduce_threshold",
      "value": 800,
      "effectOnMetric": "down"
    }
  },
  "log": {
    "taskId": "origin-202",
    "model": "origin",
    "solver": "codex-gpt-5.6-sol",
    "calls": [
      {
        "kind": "probe",
        "metric": "finalPopulation",
        "stats": {
          "metric": "finalPopulation",
          "label": "Final Population",
          "controlMean": 67.16666666666667,
          "controlCI": [
            55.416666666666664,
            78.75
          ],
          "treatmentMean": 25.083333333333332,
          "treatmentCI": [
            15.583333333333334,
            36.083333333333336
          ],
          "deltaPct": -62.65508684863524,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 2,
            "p": 0.004192722492179168
          },
          "pAdjusted": 0.020963612460895842,
          "significant": true,
          "effect": {
            "name": "rank-biserial r",
            "value": 0.8265622349255549,
            "magnitude": "large"
          }
        },
        "configA": {}
      },
      {
        "kind": "experiment",
        "metric": "finalPopulation",
        "stats": {
          "metric": "finalPopulation",
          "label": "Final Population",
          "controlMean": 67.16666666666667,
          "controlCI": [
            55.416666666666664,
            78.75
          ],
          "treatmentMean": 67,
          "treatmentCI": [
            44.833333333333336,
            87.83333333333333
          ],
          "deltaPct": -0.2481389578163842,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 39,
            "p": 0.9999999989503827
          },
          "pAdjusted": 1,
          "significant": false,
          "effect": {
            "name": "rank-biserial r",
            "value": 0,
            "magnitude": "negligible"
          }
        },
        "configA": {},
        "configB": {
          "attack_damage": 20
        }
      },
      {
        "kind": "experiment",
        "metric": "finalPopulation",
        "stats": {
          "metric": "finalPopulation",
          "label": "Final Population",
          "controlMean": 67.16666666666667,
          "controlCI": [
            55.416666666666664,
            78.75
          ],
          "treatmentMean": 75.5,
          "treatmentCI": [
            63,
            86.33333333333333
          ],
          "deltaPct": 12.40694789081885,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 24.5,
            "p": 0.2719108016043412
          },
          "pAdjusted": 1,
          "significant": false,
          "effect": {
            "name": "rank-biserial r",
            "value": -0.3171595774809396,
            "magnitude": "medium"
          }
        },
        "configA": {},
        "configB": {
          "vision_range": 1000
        }
      },
      {
        "kind": "experiment",
        "metric": "finalPopulation",
        "stats": {
          "metric": "finalPopulation",
          "label": "Final Population",
          "controlMean": 67.16666666666667,
          "controlCI": [
            55.416666666666664,
            78.75
          ],
          "treatmentMean": 26.25,
          "treatmentCI": [
            19.5,
            33.75
          ],
          "deltaPct": -60.9181141439206,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 0,
            "p": 0.0025263126954468973
          },
          "pAdjusted": 0.014984087748136865,
          "significant": true,
          "effect": {
            "name": "rank-biserial r",
            "value": 0.8718533162913388,
            "magnitude": "large"
          }
        },
        "configA": {},
        "configB": {
          "reproduce_threshold": 1000
        }
      },
      {
        "kind": "probe",
        "metric": "finalPopulation",
        "stats": {
          "metric": "finalPopulation",
          "label": "Final Population",
          "controlMean": 26.25,
          "controlCI": [
            19.333333333333332,
            34.166666666666664
          ],
          "treatmentMean": 25.083333333333332,
          "treatmentCI": [
            15.583333333333334,
            36.083333333333336
          ],
          "deltaPct": -4.444444444444449,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 32,
            "p": 0.609845035502727
          },
          "pAdjusted": 1,
          "significant": false,
          "effect": {
            "name": "rank-biserial r",
            "value": 0.1473093729572844,
            "magnitude": "small"
          }
        },
        "configA": {
          "reproduce_threshold": 1000
        }
      }
    ],
    "claims": [],
    "submission": {
      "param": "reproduce_threshold",
      "direction": "up"
    },
    "callsUsed": 5,
    "budget": 8
  },
  "score": {
    "taskId": "origin-202",
    "model": "origin",
    "solver": "codex-gpt-5.6-sol",
    "correctParam": true,
    "correctDirection": false,
    "solved": false,
    "withinBudget": true,
    "callsUsed": 5,
    "budget": 8,
    "efficiency": 0.375,
    "claimValidity": 1,
    "score": 67.5,
    "notes": [
      "wrong/no direction (truth: down)"
    ]
  },
  "audit": {
    "metric": "finalPopulation",
    "alpha": 0.05,
    "familySize": 3,
    "distinctParamsTested": 3,
    "excessTesting": false,
    "survivors": [
      "reproduce_threshold"
    ],
    "submissionBacking": "isolating-experiment",
    "backingSurvivesHolm": true,
    "pHacking": false,
    "notes": []
  },
  "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": "fd90d549028c1ae5ea80cf7e2c6f1343f450f5b5",
    "node": "v25.2.1",
    "createdAt": "2026-07-16T14:39:20.819Z"
  }
}