{
  "dataset": "BANKING77",
  "source_commit": "57ec275d8078af65b7731c2a98be812d844a6d6b",
  "seed": 20261021,
  "model": "word TF-IDF unigrams/bigrams + multinomial logistic regression",
  "tfidf": {
    "ngram_range": [
      1,
      2
    ],
    "sublinear_tf": true,
    "min_df": 2
  },
  "logistic_regression": {
    "C": 10.0,
    "max_iter": 1000,
    "solver": "lbfgs",
    "tol": 0.0001
  },
  "training_fraction": 0.7,
  "split": "Fixed random permutation of unique normalized official training texts; first 70% for training, rest calibration. Official test kept for final evaluation. Exact text duplicates removed.",
  "score": "maximum predicted class probability; used only as ranking score",
  "thresholds": [
    0.35,
    0.45,
    0.55,
    0.65,
    0.7,
    0.75,
    0.8,
    0.85,
    0.9,
    0.95
  ],
  "targets": [
    0.01,
    0.02,
    0.05,
    0.1
  ],
  "family_delta": 0.05,
  "bound": "One-sided Clopper-Pearson binomial upper bound, Bonferroni delta/10 for simultaneous threshold validity.",
  "selection": "At each target, choose threshold with most calibration selections among those with upper bound <= target; ties prefer higher threshold. No passing threshold means all requests retain review.",
  "error": "Predicted intent differs from the dataset intent label.",
  "population": "English banking support queries under the same sampling distribution; 77 known intents.",
  "scope": "Intent routing, not factual correctness of generated replies. Assumes independent examples and fixed model and threshold family. No guarantee under distribution shift, missing labels or correlated conversations.",
  "workload": "One routing review per unselected enquiry. No audits, reviewer errors, task durations, integration or setup costs modelled.",
  "tuning": "One model configuration, one split and one predeclared threshold family. No tuning on calibration/test results."
}
