{
  "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": "sonnet-4.6",
    "calls": [
      {
        "kind": "experiment",
        "metric": "finalPopulation",
        "stats": {
          "metric": "finalPopulation",
          "label": "Final Population",
          "controlMean": 67.16666666666667,
          "controlCI": [
            55.416666666666664,
            78.75
          ],
          "treatmentMean": 80.33333333333333,
          "treatmentCI": [
            69,
            92.25
          ],
          "deltaPct": 19.602977667493782,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 15,
            "p": 0.0646368510485622
          },
          "pAdjusted": 0.38782110629137323,
          "significant": false,
          "effect": {
            "name": "rank-biserial r",
            "value": -0.5334025586191666,
            "magnitude": "large"
          }
        },
        "configA": {},
        "configB": {
          "attack_damage": 2
        }
      },
      {
        "kind": "experiment",
        "metric": "finalPopulation",
        "stats": {
          "metric": "finalPopulation",
          "label": "Final Population",
          "controlMean": 67.16666666666667,
          "controlCI": [
            55.416666666666664,
            78.75
          ],
          "treatmentMean": 74.91666666666667,
          "treatmentCI": [
            61.75,
            92.58333333333333
          ],
          "deltaPct": 11.538461538461538,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 33,
            "p": 0.6661372645006556
          },
          "pAdjusted": 1,
          "significant": false,
          "effect": {
            "name": "rank-biserial r",
            "value": -0.12455047375590553,
            "magnitude": "small"
          }
        },
        "configA": {},
        "configB": {
          "mutation_rate": 2
        }
      },
      {
        "kind": "experiment",
        "metric": "finalPopulation",
        "stats": {
          "metric": "finalPopulation",
          "label": "Final Population",
          "controlMean": 67.16666666666667,
          "controlCI": [
            55.416666666666664,
            78.75
          ],
          "treatmentMean": 205.41666666666666,
          "treatmentCI": [
            144.25,
            263.6666666666667
          ],
          "deltaPct": 205.83126550868488,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 1,
            "p": 0.0032519427764485798
          },
          "pAdjusted": 0.0162597138822429,
          "significant": true,
          "effect": {
            "name": "rank-biserial r",
            "value": -0.849534582538231,
            "magnitude": "large"
          }
        },
        "configA": {},
        "configB": {
          "reproduce_threshold": 2
        }
      }
    ],
    "claims": [],
    "submission": {
      "param": "reproduce_threshold",
      "direction": "up",
      "magnitude": "large"
    },
    "callsUsed": 3,
    "budget": 8
  },
  "score": {
    "taskId": "origin-202-l2",
    "model": "origin",
    "solver": "sonnet-4.6",
    "correctParam": true,
    "correctDirection": false,
    "correctMagnitude": false,
    "solved": false,
    "withinBudget": true,
    "callsUsed": 3,
    "budget": 8,
    "efficiency": 0.625,
    "claimValidity": 1,
    "score": 69.4,
    "notes": [
      "wrong/no direction (truth: down)",
      "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": "d86631b2d89b26b1472009e69619f41647f2f1fb",
    "node": "v25.2.1",
    "createdAt": "2026-06-12T06:28:00.501Z"
  }
}