{
  "task": {
    "id": "origin-101-l2",
    "model": "origin",
    "seed": 101,
    "control": {},
    "metric": "finalPopulation",
    "replicates": 12,
    "budget": 8,
    "tier": "L2",
    "candidateParams": [
      "vision_range",
      "reproduce_threshold",
      "attack_damage"
    ],
    "truth": {
      "param": "reproduce_threshold",
      "value": 322,
      "effectOnMetric": "down",
      "magnitude": "medium"
    }
  },
  "log": {
    "taskId": "origin-101-l2",
    "model": "origin",
    "solver": "glm-5.1",
    "calls": [
      {
        "kind": "experiment",
        "metric": "finalPopulation",
        "stats": {
          "metric": "finalPopulation",
          "label": "Final Population",
          "controlMean": 74.83333333333333,
          "controlCI": [
            55.583333333333336,
            92.16666666666667
          ],
          "treatmentMean": 68.25,
          "treatmentCI": [
            48,
            88
          ],
          "deltaPct": -8.797327394209347,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 25.5,
            "p": 0.307635369736726
          },
          "pAdjusted": 0.615270739473452,
          "significant": false,
          "effect": {
            "name": "rank-biserial r",
            "value": 0.2945053219465868,
            "magnitude": "small"
          }
        },
        "configA": {
          "vision_range": 1
        },
        "configB": {
          "vision_range": 10
        }
      },
      {
        "kind": "experiment",
        "metric": "finalPopulation",
        "stats": {
          "metric": "finalPopulation",
          "label": "Final Population",
          "controlMean": 77.25,
          "controlCI": [
            67.25,
            86.66666666666667
          ],
          "treatmentMean": 73.33333333333333,
          "treatmentCI": [
            58.333333333333336,
            86.66666666666667
          ],
          "deltaPct": -5.070118662351678,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 30.5,
            "p": 0.5301262295068148
          },
          "pAdjusted": 1,
          "significant": false,
          "effect": {
            "name": "rank-biserial r",
            "value": 0.18123404427482265,
            "magnitude": "small"
          }
        },
        "configA": {
          "attack_damage": 1
        },
        "configB": {
          "attack_damage": 10
        }
      },
      {
        "kind": "experiment",
        "metric": "finalPopulation",
        "stats": {
          "metric": "finalPopulation",
          "label": "Final Population",
          "controlMean": 228.5,
          "controlCI": [
            168.5,
            281.5833333333333
          ],
          "treatmentMean": 182.83333333333334,
          "treatmentCI": [
            122.66666666666667,
            238
          ],
          "deltaPct": -19.9854121079504,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 26,
            "p": 0.3267995636020937
          },
          "pAdjusted": 0.5390023555672467,
          "significant": false,
          "effect": {
            "name": "rank-biserial r",
            "value": 0.28306925853614895,
            "magnitude": "small"
          }
        },
        "configA": {
          "reproduce_threshold": 1
        },
        "configB": {
          "reproduce_threshold": 10
        }
      },
      {
        "kind": "experiment",
        "metric": "finalPopulation",
        "stats": {
          "metric": "finalPopulation",
          "label": "Final Population",
          "controlMean": 228.5,
          "controlCI": [
            168.5,
            281.5833333333333
          ],
          "treatmentMean": 96.83333333333333,
          "treatmentCI": [
            68.83333333333333,
            123
          ],
          "deltaPct": -57.62217359591539,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 3,
            "p": 0.005355381445612206
          },
          "pAdjusted": 0.021421525782448825,
          "significant": true,
          "effect": {
            "name": "rank-biserial r",
            "value": 0.8039166942426631,
            "magnitude": "large"
          }
        },
        "configA": {
          "reproduce_threshold": 1
        },
        "configB": {
          "reproduce_threshold": 50
        }
      }
    ],
    "claims": [],
    "submission": {
      "param": "reproduce_threshold",
      "direction": "up",
      "magnitude": "medium"
    },
    "callsUsed": 4,
    "budget": 8
  },
  "score": {
    "taskId": "origin-101-l2",
    "model": "origin",
    "solver": "glm-5.1",
    "correctParam": true,
    "correctDirection": false,
    "correctMagnitude": true,
    "solved": false,
    "withinBudget": true,
    "callsUsed": 4,
    "budget": 8,
    "efficiency": 0.5,
    "claimValidity": 1,
    "score": 77.5,
    "notes": [
      "wrong/no direction (truth: down)"
    ]
  },
  "audit": {
    "metric": "finalPopulation",
    "alpha": 0.05,
    "familySize": 4,
    "distinctParamsTested": 3,
    "excessTesting": true,
    "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": "b1010caca9edeebf446c741ad0df4d13591619bf",
    "node": "v25.2.1",
    "createdAt": "2026-06-10T21:38:12.890Z"
  }
}