{
  "about": "Method-family labels from (tool id, arguments) of each agent attempt's answer call (magi_bench.labels): tool-family accuracy from tool calls, not the method the model named.",
  "labels": {
    "nist_univariate.lew.agent.e1.a1": {
      "family": "other",
      "rule": "not an estimator in the table",
      "tool_id": "data.describe"
    },
    "nist_univariate.lottery.agent.e1.a1": {
      "family": "other",
      "rule": "not an estimator in the table",
      "tool_id": "data.describe"
    },
    "nist_univariate.mavro.agent.e1.a1": {
      "family": "other",
      "rule": "not an estimator in the table",
      "tool_id": "data.describe"
    },
    "nist_univariate.michelso.agent.e1.a1": {
      "family": "other",
      "rule": "not an estimator in the table",
      "tool_id": "data.describe"
    },
    "nist_univariate.numacc1.agent.e1.a1": {
      "family": "other",
      "rule": "not an estimator in the table",
      "tool_id": "data.describe"
    },
    "nist_univariate.numacc2.agent.e1.a1": {
      "family": "other",
      "rule": "not an estimator in the table",
      "tool_id": "data.describe"
    },
    "nist_univariate.numacc3.agent.e1.a1": {
      "family": "other",
      "rule": "not an estimator in the table",
      "tool_id": "data.describe"
    },
    "nist_univariate.numacc4.agent.e1.a1": {
      "family": "other",
      "rule": "not an estimator in the table",
      "tool_id": "data.describe"
    },
    "nist_univariate.pidigits.agent.e1.a1": {
      "family": "other",
      "rule": "not an estimator in the table",
      "tool_id": "data.describe"
    }
  }
}