{
  "task": {
    "id": "social-202-l3",
    "model": "social",
    "seed": 202,
    "control": {},
    "metric": "polarization",
    "replicates": 12,
    "budget": 8,
    "tier": "L3",
    "candidateParams": [
      "mu",
      "eps",
      "meanDeg",
      "k"
    ],
    "truth": {
      "param": "meanDeg",
      "value": 3,
      "effectOnMetric": "up",
      "param2": "eps",
      "value2": 0.08,
      "interaction": "negative"
    }
  },
  "log": {
    "taskId": "social-202-l3",
    "model": "social",
    "solver": "sonnet-4.6",
    "calls": [
      {
        "kind": "experiment",
        "metric": "polarization",
        "stats": {
          "metric": "polarization",
          "label": "Polarization Index",
          "controlMean": 0.0942172593677242,
          "controlCI": [
            0.0733141981084676,
            0.11746990406221207
          ],
          "treatmentMean": 0.06848208585324757,
          "treatmentCI": [
            0.058633328995711054,
            0.07704143714108795
          ],
          "deltaPct": -27.31471249236175,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 18,
            "p": 0.10780181584888271
          },
          "pAdjusted": 0.754612710942179,
          "significant": false,
          "effect": {
            "name": "rank-biserial r",
            "value": 0.46423358399928427,
            "magnitude": "medium"
          }
        },
        "configA": {},
        "configB": {
          "mu": 0.3
        }
      },
      {
        "kind": "experiment",
        "metric": "polarization",
        "stats": {
          "metric": "polarization",
          "label": "Polarization Index",
          "controlMean": 0.0942172593677242,
          "controlCI": [
            0.0733141981084676,
            0.11746990406221207
          ],
          "treatmentMean": 0.047967018836546534,
          "treatmentCI": [
            0.03949269587295324,
            0.0562591052762097
          ],
          "deltaPct": -49.08892578870906,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 9,
            "p": 0.02065832536056611
          },
          "pAdjusted": 0.10329162680283055,
          "significant": false,
          "effect": {
            "name": "rank-biserial r",
            "value": 0.6680434501453115,
            "magnitude": "large"
          }
        },
        "configA": {},
        "configB": {
          "eps": 0.3
        }
      },
      {
        "kind": "experiment",
        "metric": "polarization",
        "stats": {
          "metric": "polarization",
          "label": "Polarization Index",
          "controlMean": 0.0942172593677242,
          "controlCI": [
            0.0733141981084676,
            0.11746990406221207
          ],
          "treatmentMean": 0.03325188739994094,
          "treatmentCI": [
            0.02545137781356364,
            0.041447520211659136
          ],
          "deltaPct": -64.7072228346604,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 0,
            "p": 0.0025263126954468973
          },
          "pAdjusted": 0.020133079879347093,
          "significant": true,
          "effect": {
            "name": "rank-biserial r",
            "value": 0.8718533162913388,
            "magnitude": "large"
          }
        },
        "configA": {},
        "configB": {
          "meanDeg": 10
        }
      },
      {
        "kind": "experiment",
        "metric": "polarization",
        "stats": {
          "metric": "polarization",
          "label": "Polarization Index",
          "controlMean": 0.0942172593677242,
          "controlCI": [
            0.0733141981084676,
            0.11746990406221207
          ],
          "treatmentMean": 0.06848208585324757,
          "treatmentCI": [
            0.058633328995711054,
            0.07704143714108795
          ],
          "deltaPct": -27.31471249236175,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 18,
            "p": 0.10780181584888271
          },
          "pAdjusted": 0.754612710942179,
          "significant": false,
          "effect": {
            "name": "rank-biserial r",
            "value": 0.46423358399928427,
            "magnitude": "medium"
          }
        },
        "configA": {},
        "configB": {
          "k": 5
        }
      },
      {
        "kind": "experiment",
        "metric": "polarization",
        "stats": {
          "metric": "polarization",
          "label": "Polarization Index",
          "controlMean": 0.0942172593677242,
          "controlCI": [
            0.0733141981084676,
            0.11746990406221207
          ],
          "treatmentMean": 0.028216764066504158,
          "treatmentCI": [
            0.020295898943246963,
            0.03739063983375672
          ],
          "deltaPct": -70.05138521767455,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 0,
            "p": 0.0025263126954468973
          },
          "pAdjusted": 0.020055840606556075,
          "significant": true,
          "effect": {
            "name": "rank-biserial r",
            "value": 0.8718533162913388,
            "magnitude": "large"
          }
        },
        "configA": {},
        "configB": {
          "meanDeg": 10,
          "eps": 0.3
        }
      },
      {
        "kind": "probe",
        "metric": "polarization",
        "stats": {
          "metric": "polarization",
          "label": "Polarization Index",
          "controlMean": 0.04039852652009964,
          "controlCI": [
            0.02571590354879542,
            0.05609147765062516
          ],
          "treatmentMean": 0.21834994723309742,
          "treatmentCI": [
            0.19918629023110648,
            0.23698388214397711
          ],
          "deltaPct": 440.489879313942,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 0,
            "p": 0.0025263126954468973
          },
          "pAdjusted": 0.02021050156357518,
          "significant": true,
          "effect": {
            "name": "rank-biserial r",
            "value": -0.8718533162913388,
            "magnitude": "large"
          }
        },
        "configA": {
          "meanDeg": 10,
          "eps": 0.3
        }
      },
      {
        "kind": "probe",
        "metric": "polarization",
        "stats": {
          "metric": "polarization",
          "label": "Polarization Index",
          "controlMean": 0.5010878587908033,
          "controlCI": [
            0.49010462261794596,
            0.5125564943523689
          ],
          "treatmentMean": 0.21834994723309742,
          "treatmentCI": [
            0.19918629023110648,
            0.23698388214397711
          ],
          "deltaPct": -56.42481784332052,
          "test": {
            "name": "Wilcoxon signed-rank",
            "statistic": 0,
            "p": 0.0025263126954468973
          },
          "pAdjusted": 0.02021050156357518,
          "significant": true,
          "effect": {
            "name": "rank-biserial r",
            "value": 0.8718533162913388,
            "magnitude": "large"
          }
        },
        "configA": {
          "eps": 0.05,
          "meanDeg": 2
        }
      }
    ],
    "claims": [],
    "submission": {
      "param": "eps",
      "param2": "meanDeg",
      "interaction": "negative"
    },
    "callsUsed": 7,
    "budget": 8
  },
  "score": {
    "taskId": "social-202-l3",
    "model": "social",
    "solver": "sonnet-4.6",
    "correctParam": true,
    "correctDirection": false,
    "correctParams": true,
    "correctInteraction": true,
    "solved": true,
    "withinBudget": true,
    "callsUsed": 7,
    "budget": 8,
    "efficiency": 0.125,
    "claimValidity": 1,
    "score": 57.5,
    "notes": [
      "not backed by a 2x2 factorial (isolate each param + a combined experiment)"
    ]
  },
  "audit": {
    "metric": "polarization",
    "alpha": 0.05,
    "familySize": 4,
    "distinctParamsTested": 4,
    "excessTesting": false,
    "survivors": [
      "meanDeg"
    ],
    "submissionBacking": "isolating-experiment",
    "backingSurvivesHolm": false,
    "pHacking": false,
    "notes": [
      "'eps' is significant alone but borderline: it does not survive family-wide Holm (low power, not over-testing)"
    ]
  },
  "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": "2026-06-11T15:42:00.240Z"
  }
}