{
 "generated": "2026-09-27",
 "reranker": "Laya (Ollaya 0.7.3, open weights), CPU",
 "rule": "When Decide's top probability < tau, Laya chooses among Decide's top 5 labels; p = (1-w)*Decide renormalised over the top 5 + w*Laya.",
 "datasets": [
  {
   "dataset": "banking77",
   "n_test": 3080,
   "decide_accuracy_pct": 91.56,
   "dev_chosen": {
    "tau": 0.5,
    "w": 0.3,
    "accuracy_pct": 91.79,
    "sent_to_laya": 309,
    "fixed": 27,
    "broke": 20,
    "mcnemar_exact_p": 0.38
   },
   "service_default": {
    "tau": 0.7,
    "w": 0.3,
    "accuracy_pct": 91.82,
    "sent_to_laya": 599,
    "fixed": 31,
    "broke": 23,
    "mcnemar_exact_p": 0.34
   },
   "laya_ms_p50_eval_box": 1605
  },
  {
   "dataset": "hwu64",
   "n_test": 1076,
   "decide_accuracy_pct": 87.45,
   "dev_chosen": {
    "tau": 0.6,
    "w": 0.4,
    "accuracy_pct": 88.38,
    "sent_to_laya": 220,
    "fixed": 29,
    "broke": 19,
    "mcnemar_exact_p": 0.19
   },
   "service_default": {
    "tau": 0.7,
    "w": 0.3,
    "accuracy_pct": 88.66,
    "sent_to_laya": 295,
    "fixed": 25,
    "broke": 12,
    "mcnemar_exact_p": 0.047
   },
   "laya_ms_p50_eval_box": 948
  },
  {
   "dataset": "clinc150",
   "n_test": 4500,
   "decide_accuracy_pct": 93.0,
   "dev_chosen": {
    "tau": 0.7,
    "w": 0.4,
    "accuracy_pct": 94.67,
    "sent_to_laya": 862,
    "fixed": 119,
    "broke": 44,
    "mcnemar_exact_p": 3.6e-09
   },
   "service_default": {
    "tau": 0.7,
    "w": 0.3,
    "accuracy_pct": 94.33,
    "sent_to_laya": 862,
    "fixed": 87,
    "broke": 27,
    "mcnemar_exact_p": 1.5e-08
   },
   "laya_ms_p50_eval_box": 1146
  }
 ],
 "pooled": {
  "dev_chosen": {
   "fixed": 175,
   "broke": 83,
   "mcnemar_exact_p": 1e-08
  },
  "service_default": {
   "fixed": 143,
   "broke": 62,
   "mcnemar_exact_p": 1.5e-08
  }
 }
}