{
  "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": "ofat-rand",
    "calls": [
      {
        "kind": "experiment",
        "metric": "finalPopulation",
        "stats": {
          "metric": "finalPopulation",
          "label": "Final Population",
          "controlMean": 67.16666666666667,
          "controlCI": [
            55.416666666666664,
            78.75
          ],
          "treatmentMean": 59.25,
          "treatmentCI": [
            44.916666666666664,
            74
          ],
          "deltaPct": -11.786600496277922,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 23,
            "p": 0.22383729098503746
          },
          "pAdjusted": 1,
          "significant": false,
          "effect": {
            "name": "rank-biserial r",
            "value": 0.35114096078246887,
            "magnitude": "medium"
          }
        },
        "configA": {},
        "configB": {
          "attack_damage": 81.91878627985716
        }
      },
      {
        "kind": "experiment",
        "metric": "finalPopulation",
        "stats": {
          "metric": "finalPopulation",
          "label": "Final Population",
          "controlMean": 67.16666666666667,
          "controlCI": [
            55.416666666666664,
            78.75
          ],
          "treatmentMean": 69.16666666666667,
          "treatmentCI": [
            53.583333333333336,
            82.66666666666667
          ],
          "deltaPct": 2.9776674937965257,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 35,
            "p": 0.7835712448792027
          },
          "pAdjusted": 1,
          "significant": false,
          "effect": {
            "name": "rank-biserial r",
            "value": -0.0792898943702349,
            "magnitude": "negligible"
          }
        },
        "configA": {},
        "configB": {
          "mutation_rate": 0.1048421838786453
        }
      },
      {
        "kind": "experiment",
        "metric": "finalPopulation",
        "stats": {
          "metric": "finalPopulation",
          "label": "Final Population",
          "controlMean": 67.16666666666667,
          "controlCI": [
            55.416666666666664,
            78.75
          ],
          "treatmentMean": 37,
          "treatmentCI": [
            25.916666666666668,
            49.166666666666664
          ],
          "deltaPct": -44.91315136476427,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 8,
            "p": 0.01668675312635237
          },
          "pAdjusted": 0.08343376563176186,
          "significant": false,
          "effect": {
            "name": "rank-biserial r",
            "value": 0.6909547937977613,
            "magnitude": "large"
          }
        },
        "configA": {},
        "configB": {
          "reproduce_threshold": 699.6860284861177
        }
      }
    ],
    "claims": [
      {
        "param": "attack_damage",
        "effect": "none"
      },
      {
        "param": "mutation_rate",
        "effect": "none"
      },
      {
        "param": "reproduce_threshold",
        "effect": "none"
      }
    ],
    "submission": null,
    "callsUsed": 3,
    "budget": 8
  },
  "score": {
    "taskId": "origin-202-l2",
    "model": "origin",
    "solver": "ofat-rand",
    "correctParam": false,
    "correctDirection": false,
    "correctMagnitude": false,
    "solved": false,
    "withinBudget": true,
    "callsUsed": 3,
    "budget": 8,
    "efficiency": 0.625,
    "claimValidity": 1,
    "score": 9.4,
    "notes": [
      "wrong/no param (truth: reproduce_threshold)",
      "wrong/no direction (truth: down)",
      "no magnitude submitted (truth: medium)",
      "conclusion not backed by a significant isolating experiment"
    ]
  },
  "audit": {
    "metric": "finalPopulation",
    "alpha": 0.05,
    "familySize": 3,
    "distinctParamsTested": 3,
    "excessTesting": false,
    "survivors": [],
    "submissionBacking": "none",
    "backingSurvivesHolm": null,
    "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": "9a93d577a9e675a4ef200b7eaf8d0507b7fde39a",
    "node": "v25.2.1",
    "createdAt": "2026-06-12T17:33:16.238Z"
  }
}