{
  "experiment": {
    "threshold": 0.9903820753097534,
    "val": {
      "recall_at_cal_thr": 0.9677236693091733,
      "wilson": [
        0.9584121604850137,
        0.9750047057475352
      ],
      "fpr_achieved": 0.010020876826722338,
      "oracle_recall_fpr1": 0.9677236693091733,
      "oracle_wilson": [
        0.9584121604850137,
        0.9750047057475352
      ],
      "solo_oracle": {
        "m1_polish_roberta": 0.9490373725934315,
        "m2_xlm_roberta": 0.964326160815402
      },
      "n_ai": 1766,
      "n_human": 2395
    },
    "test_a": {
      "recall_at_cal_thr": 0.905977945443993,
      "wilson": [
        0.8977171084066199,
        0.913635785136457
      ],
      "fpr_achieved": 0.0991495941244685,
      "oracle_recall_fpr1": 0.591603791835945,
      "oracle_wilson": [
        0.5781404327776718,
        0.6049310921652145
      ],
      "solo_oracle": {
        "m1_polish_roberta": 0.5925710969239698,
        "m2_xlm_roberta": 0.600309537628168
      },
      "n_ai": 5169,
      "n_human": 5174
    },
    "test_b": {
      "recall_at_cal_thr": 0.9061548499347543,
      "wilson": [
        0.9000238428625889,
        0.9119466588903978
      ],
      "fpr_achieved": 0.10697271546123863,
      "oracle_recall_fpr1": 0.6358199217050892,
      "oracle_wilson": [
        0.6259299299428169,
        0.6455964841641391
      ],
      "solo_oracle": {
        "m1_polish_roberta": 0.5892779469334494,
        "m2_xlm_roberta": 0.604066985645933
      },
      "n_ai": 9196,
      "n_human": 9236
    },
    "anti_shortcut": {
      "n": 506,
      "acc@0.5": 0.9861660079051383,
      "recall@thr": 1.0,
      "fpr@thr": 0.0
    },
    "humanized_recall": 1.0
  },
  "int8_calibration": {
    "threshold": 0.5046048760414124,
    "val_recall": 0.9603624009060022,
    "val_fpr": 0.010020876826722338
  },
  "warning": "Historical experiment and calibration, not a new independent production benchmark."
}