{
 "score_key": "p_true_think",
 "gate": {
  "pool_n": 2130,
  "pool_pass_rate": 0.033,
  "pool_auroc_pooled": 0.961,
  "pool_auroc_within_mean": 0.845,
  "n_mixed_tasks": 20,
  "n_pairs": 70,
  "score_by_depth": {
   "1": 0.363,
   "2": 0.097,
   "3": 0.069
  },
  "quota_conf_strat": {
   "1": 15,
   "2": 30,
   "3": 25
  },
  "exec_conf_strat_jaccard": 0.273
 },
 "exec": {
  "n_pairs": 70,
  "purity_full_pass": 1.0,
  "visible_pass_rate": 1.0,
  "distinct_tasks": 31,
  "depth_mix": {
   "1": 43,
   "2": 20,
   "3": 7
  },
  "mean_score": 0.611,
  "min_score_kept": 0.014,
  "mean_p_true_nothink": 0.476,
  "mean_freq": 10.7
 },
 "conf_strat": {
  "n_pairs": 70,
  "purity_full_pass": 0.429,
  "visible_pass_rate": 0.429,
  "distinct_tasks": 31,
  "depth_mix": {
   "1": 15,
   "2": 30,
   "3": 25
  },
  "mean_score": 0.69,
  "min_score_kept": 0.392,
  "mean_p_true_nothink": 0.632,
  "mean_freq": 6.3
 },
 "conf_global": {
  "n_pairs": 70,
  "purity_full_pass": 0.614,
  "visible_pass_rate": 0.614,
  "distinct_tasks": 31,
  "depth_mix": {
   "1": 36,
   "2": 29,
   "3": 5
  },
  "mean_score": 0.753,
  "min_score_kept": 0.523,
  "mean_p_true_nothink": 0.641,
  "mean_freq": 8.9
 },
 "rand": {
  "n_pairs": 70,
  "purity_full_pass": 0.029,
  "visible_pass_rate": 0.029,
  "distinct_tasks": 42,
  "depth_mix": {
   "3": 34,
   "2": 29,
   "1": 7
  },
  "mean_score": 0.107,
  "min_score_kept": 0.0,
  "mean_p_true_nothink": 0.235,
  "mean_freq": 1.0
 }
}