{
  "calibration": "Separate NLL temperature fit for base and adapter: 81 log-spaced values in [0.25,4]",
  "historical_baseline": "Same training sample; airline/origin/4-hour departure bin; 20-observation global prior; global Beta(1,1) smoothing",
  "input_fields": [
    "flight_date",
    "day_of_week",
    "airline",
    "flight_number",
    "origin",
    "destination",
    "scheduled_departure_local",
    "scheduled_arrival_local",
    "scheduled_duration_minutes",
    "distance_miles"
  ],
  "months_2025": {
    "calibration": 2,
    "test": 3,
    "train": 1
  },
  "origins": [
    "EWR",
    "JFK",
    "LGA"
  ],
  "population": "Completed, non-diverted NYC departures with observed arrival delay",
  "primary_metric": "Mean binary Brier: mean((P(late)-observed_late)^2); lower is better",
  "promotion": "Research only; no automatic model registration or serving changes",
  "question": {
    "criteria": {
      "late": "Arrives at least 15 minutes late.",
      "not_late": "Arrives less than 15 minutes late, including early or on time."
    },
    "instructions": "Estimate whether this flight will arrive at least 15 minutes late. Only scheduled information is available before departure. This benchmark covers completed, non-diverted flights with recorded arrival times.",
    "type": "choice"
  },
  "sample_counts": {
    "calibration": 512,
    "test": 1024,
    "train": 3072
  },
  "sampling": "Smallest SHA256(seed:flight-identity), no label balancing",
  "seed": 553,
  "success": "Adapter beats equally calibrated base and historical baseline on test Brier; both paired date-bootstrap 95% intervals exclude zero",
  "training": {
    "alpha": 32,
    "dropout": 0.05,
    "epochs": 1,
    "gradient_accumulation": 4,
    "learning_rate": 2e-05,
    "rank": 16,
    "seed": 553
  },
  "uncertainty": "2000 paired bootstrap resamples of whole test dates, seed 553",
  "version": "zils-flight-delay-001"
}