{
 "prompt": 2,
 "orders": 1,
 "method": "temperature per question type and per option-count bucket (choice), by NLL, fitted on the out-of-distribution calibration split only; thresholds per type: the most permissive threshold whose accepted calibration decisions have a one-sided 95% Clopper-Pearson upper bound <= max_error; a type without one defers",
 "temperature_per_prim": {
  "choice": 1.2599210498948732,
  "noul": 1.8340080864093422,
  "score": 1.9713263972803747
 },
 "temperature_per_bucket": {
  "choice": {
   "2-10": 1.6339154532410998,
   "11-26": 0.9576032806985735,
   "27-255": 0.8908987181403392
  }
 },
 "thresholds": {
  "0.01": {
   "confidence": null,
   "method": "per type: the most permissive threshold whose accepted calibration decisions have a one-sided 95% Clopper-Pearson upper bound <= max_error; a type without one defers",
   "per_type": {
    "choice": {
     "confidence": null,
     "heldout_coverage": 0.0,
     "heldout_error": null
    },
    "noul": {
     "confidence": null,
     "heldout_coverage": 0.0,
     "heldout_error": null
    },
    "score": {
     "confidence": null,
     "heldout_coverage": 0.0,
     "heldout_error": null
    }
   },
   "heldout_coverage": 0.0,
   "heldout_error": null
  },
  "0.05": {
   "confidence": null,
   "method": "per type: the most permissive threshold whose accepted calibration decisions have a one-sided 95% Clopper-Pearson upper bound <= max_error; a type without one defers",
   "per_type": {
    "choice": {
     "confidence": 0.9794658159140632,
     "heldout_coverage": 0.07419572174498905,
     "heldout_error": 0.009080590238365494
    },
    "noul": {
     "confidence": 0.9139974126226222,
     "heldout_coverage": 0.13509933774834437,
     "heldout_error": 0.03501400560224088
    },
    "score": {
     "confidence": null,
     "heldout_coverage": 0.0,
     "heldout_error": null
    }
   },
   "heldout_coverage": 0.07920742911059245,
   "heldout_error": 0.020689655172413834
  }
 },
 "gates": {
  "ece_lofo": {
   "choice": {
    "pooled": 0.023511479853082662,
    "n": 11874,
    "gate": 0.05,
    "pass": true
   },
   "noul": {
    "pooled": 0.027070025115957533,
    "n": 5285,
    "gate": 0.05,
    "pass": true
   },
   "score": {
    "pooled": 0.027462710825686312,
    "n": 2978,
    "gate": 0.05,
    "pass": true
   }
  },
  "ece_worst_fold": 0.10555608237660483,
  "worst_fold": "agent",
  "automated_at": {
   "0.01": 0.0,
   "0.05": 0.07920742911059245
  },
  "automated_error": {
   "0.01": null,
   "0.05": 0.020689655172413834
  },
  "note": "held-out numbers: every evaluation row predicted by the fit without its family"
 },
 "test": {
  "T": {
   "choice": {
    "acc": 0.6942058278591882,
    "ece": 0.022368741135958967,
    "n": 11874
   },
   "noul": {
    "acc": 0.7458845789971618,
    "ece": 0.02910514429567303,
    "n": 5285
   },
   "score": {
    "acc": 0.5218267293485561,
    "ece": 0.027385410166595566,
    "n": 2978
   },
   "all": {
    "acc": 0.6822764066146894,
    "ece": 0.017802099792149573,
    "n": 20137
   }
  },
  "note": "evaluation pool (development and test splits), full fit"
 },
 "n": {
  "fit": 3560,
  "many": 1500,
  "eval": 20137
 },
 "engine": "fast",
 "fitted_on": {
  "split": "out-of-distribution calibration split (sealed-seed BH-, JDI- and agent-shaped families, many-option items)",
  "orders": 1,
  "readouts": "basal-serve --mode fast, ModelOpt FP8 package (recipe b2)"
 },
 "certified_on": "the same split",
 "evaluated_on": "development and test splits, leave-one-family-out and leave-one-language-out"
}