{
  "task": {
    "id": "market-101-l3",
    "model": "market",
    "seed": 101,
    "control": {},
    "metric": "volatility",
    "replicates": 12,
    "budget": 8,
    "tier": "L3",
    "candidateParams": [
      "fundFrac",
      "lambda",
      "chartFrac",
      "sigmaF"
    ],
    "truth": {
      "param": "sigmaF",
      "value": 0.006,
      "effectOnMetric": "up",
      "param2": "chartFrac",
      "value2": 0.6,
      "interaction": "negative"
    }
  },
  "log": {
    "taskId": "market-101-l3",
    "model": "market",
    "solver": "gemini-3.1-pro",
    "calls": [
      {
        "kind": "experiment",
        "metric": "volatility",
        "stats": {
          "metric": "volatility",
          "label": "Volatility (return std)",
          "controlMean": 0.002327212795344664,
          "controlCI": [
            0.00212110742919089,
            0.0025333181614984384
          ],
          "treatmentMean": 0.0032975015373722365,
          "treatmentCI": [
            0.0031989031554426435,
            0.0033960999193018296
          ],
          "deltaPct": 41.69316806647546,
          "test": {
            "name": "Mann-Whitney U (rank-sum)",
            "statistic": 0,
            "p": 0.00003660522347836981
          },
          "pAdjusted": 0.0006410927303943659,
          "significant": true,
          "effect": {
            "name": "Cliff's delta",
            "value": -1,
            "magnitude": "large"
          }
        },
        "configA": {},
        "configB": {
          "fundFrac": 0.1
        }
      },
      {
        "kind": "probe",
        "metric": "volatility",
        "stats": {
          "metric": "volatility",
          "label": "Volatility (return std)",
          "controlMean": 0.002327212795344664,
          "controlCI": [
            0.00212110742919089,
            0.0025333181614984384
          ],
          "treatmentMean": 0.11231408079054171,
          "treatmentCI": [
            0.08076031146389105,
            0.1438678501171924
          ],
          "deltaPct": 4726.1199412109545,
          "test": {
            "name": "Mann-Whitney U (rank-sum)",
            "statistic": 0,
            "p": 0.00003660522347836981
          },
          "pAdjusted": 0.0005811279681218195,
          "significant": true,
          "effect": {
            "name": "Cliff's delta",
            "value": -1,
            "magnitude": "large"
          }
        },
        "configA": {}
      },
      {
        "kind": "experiment",
        "metric": "volatility",
        "stats": {
          "metric": "volatility",
          "label": "Volatility (return std)",
          "controlMean": 0.002327212795344664,
          "controlCI": [
            0.00212110742919089,
            0.0025333181614984384
          ],
          "treatmentMean": 0.002327212795344664,
          "treatmentCI": [
            0.00212110742919089,
            0.0025333181614984384
          ],
          "deltaPct": 0,
          "test": {
            "name": "Mann-Whitney U (rank-sum)",
            "statistic": 72,
            "p": 1
          },
          "pAdjusted": 1,
          "significant": false,
          "effect": {
            "name": "Cliff's delta",
            "value": 0,
            "magnitude": "negligible"
          }
        },
        "configA": {},
        "configB": {
          "lambda": 0.1
        }
      },
      {
        "kind": "experiment",
        "metric": "volatility",
        "stats": {
          "metric": "volatility",
          "label": "Volatility (return std)",
          "controlMean": 0.002327212795344664,
          "controlCI": [
            0.00212110742919089,
            0.0025333181614984384
          ],
          "treatmentMean": 0.002327212795344664,
          "treatmentCI": [
            0.00212110742919089,
            0.0025333181614984384
          ],
          "deltaPct": 0,
          "test": {
            "name": "Mann-Whitney U (rank-sum)",
            "statistic": 72,
            "p": 1
          },
          "pAdjusted": 1,
          "significant": false,
          "effect": {
            "name": "Cliff's delta",
            "value": 0,
            "magnitude": "negligible"
          }
        },
        "configA": {},
        "configB": {
          "chartFrac": 0.1
        }
      },
      {
        "kind": "experiment",
        "metric": "volatility",
        "stats": {
          "metric": "volatility",
          "label": "Volatility (return std)",
          "controlMean": 0.002327212795344664,
          "controlCI": [
            0.00212110742919089,
            0.0025333181614984384
          ],
          "treatmentMean": 0.11231408079054171,
          "treatmentCI": [
            0.08076031146389105,
            0.1438678501171924
          ],
          "deltaPct": 4726.1199412109545,
          "test": {
            "name": "Mann-Whitney U (rank-sum)",
            "statistic": 0,
            "p": 0.00003660522347836981
          },
          "pAdjusted": 0.0005811279681218195,
          "significant": true,
          "effect": {
            "name": "Cliff's delta",
            "value": -1,
            "magnitude": "large"
          }
        },
        "configA": {},
        "configB": {
          "sigmaF": 0.1
        }
      },
      {
        "kind": "probe",
        "metric": "volatility",
        "stats": {
          "metric": "volatility",
          "label": "Volatility (return std)",
          "controlMean": 0.012008314184361343,
          "controlCI": [
            0.009829115060058536,
            0.01418751330866415
          ],
          "treatmentMean": 0.11231408079054171,
          "treatmentCI": [
            0.08076031146389105,
            0.1438678501171924
          ],
          "deltaPct": 835.3026500323459,
          "test": {
            "name": "Mann-Whitney U (rank-sum)",
            "statistic": 0,
            "p": 0.00003660522347836981
          },
          "pAdjusted": 0.000653768964137047,
          "significant": true,
          "effect": {
            "name": "Cliff's delta",
            "value": -1,
            "magnitude": "large"
          }
        },
        "configA": {
          "fundFrac": 0.1,
          "sigmaF": 0.1
        }
      },
      {
        "kind": "probe",
        "metric": "volatility",
        "stats": {
          "metric": "volatility",
          "label": "Volatility (return std)",
          "controlMean": 0.11231408079054171,
          "controlCI": [
            0.08076031146389105,
            0.1438678501171924
          ],
          "treatmentMean": 0.11231408079054171,
          "treatmentCI": [
            0.08076031146389105,
            0.1438678501171924
          ],
          "deltaPct": 0,
          "test": {
            "name": "Mann-Whitney U (rank-sum)",
            "statistic": 72,
            "p": 1
          },
          "pAdjusted": 1,
          "significant": false,
          "effect": {
            "name": "Cliff's delta",
            "value": 0,
            "magnitude": "negligible"
          }
        },
        "configA": {
          "sigmaF": 0.1,
          "lambda": 0.1
        }
      }
    ],
    "claims": [],
    "submission": {
      "param": "sigmaF",
      "param2": "lambda",
      "interaction": "negative"
    },
    "callsUsed": 7,
    "budget": 8
  },
  "score": {
    "taskId": "market-101-l3",
    "model": "market",
    "solver": "gemini-3.1-pro",
    "correctParam": true,
    "correctDirection": false,
    "correctParams": false,
    "correctInteraction": true,
    "solved": false,
    "withinBudget": true,
    "callsUsed": 7,
    "budget": 8,
    "efficiency": 0.125,
    "claimValidity": 1,
    "score": 39.5,
    "notes": [
      "only one of two params (truth: sigmaF + chartFrac)",
      "not backed by a 2x2 factorial (isolate each param + a combined experiment)"
    ]
  },
  "audit": {
    "metric": "volatility",
    "alpha": 0.05,
    "familySize": 4,
    "distinctParamsTested": 4,
    "excessTesting": false,
    "survivors": [
      "fundFrac",
      "sigmaF"
    ],
    "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": "b1010caca9edeebf446c741ad0df4d13591619bf",
    "node": "v25.2.1",
    "createdAt": "2026-06-10T19:58:17.544Z"
  }
}