{
  "task": {
    "id": "origin-202-l2",
    "model": "origin",
    "seed": 202,
    "control": {},
    "metric": "finalPopulation",
    "replicates": 12,
    "budget": 8,
    "tier": "L2",
    "candidateParams": [
      "attack_damage",
      "mutation_rate",
      "reproduce_threshold"
    ],
    "truth": {
      "param": "reproduce_threshold",
      "value": 586,
      "effectOnMetric": "down",
      "magnitude": "medium"
    }
  },
  "log": {
    "taskId": "origin-202-l2",
    "model": "origin",
    "solver": "haiku-4.5",
    "calls": [
      {
        "kind": "experiment",
        "metric": "finalPopulation",
        "stats": {
          "metric": "finalPopulation",
          "label": "Final Population",
          "controlMean": 67.16666666666667,
          "controlCI": [
            55.416666666666664,
            78.75
          ],
          "treatmentMean": 66.91666666666667,
          "treatmentCI": [
            50.833333333333336,
            81.83333333333333
          ],
          "deltaPct": -0.3722084367245657,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 37.5,
            "p": 0.9374247302379239
          },
          "pAdjusted": 1,
          "significant": false,
          "effect": {
            "name": "rank-biserial r",
            "value": -0.022662980454966834,
            "magnitude": "negligible"
          }
        },
        "configA": {},
        "configB": {
          "attack_damage": 1.5
        }
      },
      {
        "kind": "experiment",
        "metric": "finalPopulation",
        "stats": {
          "metric": "finalPopulation",
          "label": "Final Population",
          "controlMean": 67.16666666666667,
          "controlCI": [
            55.416666666666664,
            78.75
          ],
          "treatmentMean": 70,
          "treatmentCI": [
            51.166666666666664,
            86.83333333333333
          ],
          "deltaPct": 4.218362282878404,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 29,
            "p": 0.4559523774727896
          },
          "pAdjusted": 1,
          "significant": false,
          "effect": {
            "name": "rank-biserial r",
            "value": -0.21521542757635187,
            "magnitude": "small"
          }
        },
        "configA": {},
        "configB": {
          "mutation_rate": 0.15
        }
      },
      {
        "kind": "experiment",
        "metric": "finalPopulation",
        "stats": {
          "metric": "finalPopulation",
          "label": "Final Population",
          "controlMean": 67.16666666666667,
          "controlCI": [
            55.416666666666664,
            78.75
          ],
          "treatmentMean": 228.5,
          "treatmentCI": [
            169.25,
            280.8333333333333
          ],
          "deltaPct": 240.19851116625307,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 2,
            "p": 0.004192722492179168
          },
          "pAdjusted": 0.01642323032987303,
          "significant": true,
          "effect": {
            "name": "rank-biserial r",
            "value": -0.8265622349255549,
            "magnitude": "large"
          }
        },
        "configA": {},
        "configB": {
          "reproduce_threshold": 0.5
        }
      }
    ],
    "claims": [],
    "submission": {
      "param": "reproduce_threshold",
      "direction": "down",
      "magnitude": "large"
    },
    "callsUsed": 3,
    "budget": 8
  },
  "score": {
    "taskId": "origin-202-l2",
    "model": "origin",
    "solver": "haiku-4.5",
    "correctParam": true,
    "correctDirection": true,
    "correctMagnitude": false,
    "solved": false,
    "withinBudget": true,
    "callsUsed": 3,
    "budget": 8,
    "efficiency": 0.625,
    "claimValidity": 1,
    "score": 84.4,
    "notes": [
      "magnitude off by one (said large, truth medium)"
    ]
  },
  "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": "f305d9d54670c03a92b2c93b76040e15a5bc1840",
    "node": "v25.2.1",
    "createdAt": "2026-06-11T13:30:53.666Z"
  }
}