{
  "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": "gemini-3.1-pro",
    "calls": [
      {
        "kind": "probe",
        "metric": "finalPopulation",
        "stats": {
          "metric": "finalPopulation",
          "label": "Final Population",
          "controlMean": 67.16666666666667,
          "controlCI": [
            55.416666666666664,
            78.75
          ],
          "treatmentMean": 34.083333333333336,
          "treatmentCI": [
            22.25,
            44.166666666666664
          ],
          "deltaPct": -49.25558312655087,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 0,
            "p": 0.0025166349849183867
          },
          "pAdjusted": 0.012438693106627818,
          "significant": true,
          "effect": {
            "name": "rank-biserial r",
            "value": 0.8721888380725838,
            "magnitude": "large"
          }
        },
        "configA": {}
      },
      {
        "kind": "probe",
        "metric": "finalPopulation",
        "stats": {
          "metric": "finalPopulation",
          "label": "Final Population",
          "controlMean": 73.33333333333333,
          "controlCI": [
            58.5,
            87.25
          ],
          "treatmentMean": 34.083333333333336,
          "treatmentCI": [
            22.25,
            44.166666666666664
          ],
          "deltaPct": -53.522727272727266,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 0,
            "p": 0.0025166349849183867
          },
          "pAdjusted": 0.012583174924591933,
          "significant": true,
          "effect": {
            "name": "rank-biserial r",
            "value": 0.8721888380725838,
            "magnitude": "large"
          }
        },
        "configA": {
          "attack_damage": 10
        }
      },
      {
        "kind": "probe",
        "metric": "finalPopulation",
        "stats": {
          "metric": "finalPopulation",
          "label": "Final Population",
          "controlMean": 44.666666666666664,
          "controlCI": [
            29.833333333333332,
            58
          ],
          "treatmentMean": 34.083333333333336,
          "treatmentCI": [
            22.25,
            44.166666666666664
          ],
          "deltaPct": -23.694029850746258,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 22,
            "p": 0.19519505914473645
          },
          "pAdjusted": 0.7807802365789458,
          "significant": false,
          "effect": {
            "name": "rank-biserial r",
            "value": 0.3739391775069527,
            "magnitude": "medium"
          }
        },
        "configA": {
          "reproduce_threshold": 200
        }
      }
    ],
    "claims": [],
    "submission": {
      "param": "reproduce_threshold",
      "direction": "down",
      "magnitude": "medium"
    },
    "callsUsed": 3,
    "budget": 8
  },
  "score": {
    "taskId": "origin-101-l2",
    "model": "origin",
    "solver": "gemini-3.1-pro",
    "correctParam": true,
    "correctDirection": true,
    "correctMagnitude": true,
    "solved": true,
    "withinBudget": true,
    "callsUsed": 3,
    "budget": 8,
    "efficiency": 0.625,
    "claimValidity": 1,
    "score": 69.4,
    "notes": [
      "conclusion not backed by a significant isolating experiment"
    ]
  },
  "audit": {
    "metric": "finalPopulation",
    "alpha": 0.05,
    "familySize": 0,
    "distinctParamsTested": 0,
    "excessTesting": false,
    "survivors": [],
    "submissionBacking": "probe-only",
    "backingSurvivesHolm": null,
    "pHacking": false,
    "notes": [
      "submission rests on probe output-matching, not an isolating controlled experiment"
    ]
  },
  "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": "2026-06-10T19:36:18.295Z"
  }
}