{
  "overall": {
    "solo_pred": 0.5646,
    "naive_pred": 0.5207,
    "trim_pred": 0.5444,
    "median_pred": 0.5561,
    "rep_pred": 0.5172
  },
  "by_domain": {
    "airline": {
      "solo_pred": 0.3167,
      "naive_pred": 0.295,
      "trim_pred": 0.305,
      "median_pred": 0.3233,
      "rep_pred": 0.295
    },
    "retail": {
      "solo_pred": 0.6725,
      "naive_pred": 0.6188,
      "trim_pred": 0.6486,
      "median_pred": 0.6572,
      "rep_pred": 0.6138
    }
  },
  "n_trials_evaluated": 1980,
  "n_unique_tasks": 165,
  "statistical_analysis": {
    "wilson_ci": {
      "solo_pred": {
        "accuracy": 0.5646,
        "ci_lo": 0.5427,
        "ci_hi": 0.5863,
        "n": 1980
      },
      "naive_pred": {
        "accuracy": 0.5207,
        "ci_lo": 0.4987,
        "ci_hi": 0.5427,
        "n": 1980
      },
      "trim_pred": {
        "accuracy": 0.5444,
        "ci_lo": 0.5224,
        "ci_hi": 0.5663,
        "n": 1980
      },
      "median_pred": {
        "accuracy": 0.5561,
        "ci_lo": 0.5341,
        "ci_hi": 0.5778,
        "n": 1980
      },
      "rep_pred": {
        "accuracy": 0.5172,
        "ci_lo": 0.4951,
        "ci_hi": 0.5391,
        "n": 1980
      }
    },
    "task_clustered_bootstrap_ci": {
      "solo_pred": {
        "point_estimate": 0.5646464646464646,
        "boot_mean": 0.5646064141414141,
        "ci_lo": 0.5161616161616162,
        "ci_hi": 0.6126262626262626,
        "n_tasks": 165,
        "n_boot": 10000
      },
      "naive_pred": {
        "point_estimate": 0.5207070707070707,
        "boot_mean": 0.5206724242424243,
        "ci_lo": 0.4666666666666667,
        "ci_hi": 0.5737373737373738,
        "n_tasks": 165,
        "n_boot": 10000
      },
      "trim_pred": {
        "point_estimate": 0.5444444444444444,
        "boot_mean": 0.5445278282828283,
        "ci_lo": 0.49190656565656565,
        "ci_hi": 0.5964772727272726,
        "n_tasks": 165,
        "n_boot": 10000
      },
      "median_pred": {
        "point_estimate": 0.556060606060606,
        "boot_mean": 0.5560997474747474,
        "ci_lo": 0.5035353535353535,
        "ci_hi": 0.6090909090909091,
        "n_tasks": 165,
        "n_boot": 10000
      },
      "rep_pred": {
        "point_estimate": 0.5171717171717172,
        "boot_mean": 0.5171274242424242,
        "ci_lo": 0.4636363636363636,
        "ci_hi": 0.5696969696969697,
        "n_tasks": 165,
        "n_boot": 10000
      }
    },
    "mcnemar_solo_vs_aggregation": {
      "solo_vs_naive_pred": {
        "n10": 186,
        "n01": 99,
        "n_discordant": 285,
        "p_value": 2.856939193225961e-07
      },
      "solo_vs_trim_pred": {
        "n10": 139,
        "n01": 99,
        "n_discordant": 238,
        "p_value": 0.011315217519235584
      },
      "solo_vs_median_pred": {
        "n10": 117,
        "n01": 100,
        "n_discordant": 217,
        "p_value": 0.277380158804846
      },
      "solo_vs_rep_pred": {
        "n10": 200,
        "n01": 106,
        "n_discordant": 306,
        "p_value": 8.446843331902766e-08
      }
    }
  }
}