# Verification configuration. Headline results use ONLY cross-validated,
# out-of-fold predictions. Everything lead/week-specific, spatially explicit,
# characteristic-specific, probabilistic where relevant, episode-aware.

headline_weeks: [W1, W2, W3, W4, W5, W6]     # W7 supplementary, never pooled

forecast_episode_binarisation:               # GATE ITEM: confirm the rule
  proposed_rule: "daily P(EHD) >= 0.5"       # applied identically to every model
  fixed_reference_prob: 0.5
  pre_lt0_context: "observations available at initialization"
  post_lt45_buffer: "observed labels only"

metrics:
  continuous: [bias, MAE, RMSE, pearson, spearman]   # Tx, EHD_COUNT, HW_DAYS, SEV_WEEK
  count_distribution: [RPS, RPSS]                     # vs climatology; not only E[N] error
  categorical: [POD, POFD, FAR, FBI, CSI, ETS, HSS, TSS_PSS, EDI, SEDI]
  categorical_store: [H, M, FA, CN, totals, n_obs_events, n_fcst_events]
  probabilistic: [brier, BSS, ROC_AUC, PR_AUC, reliability, calibration_curve, sharpness]
  finite_sample_correction: "documented correction for POD/POFD of 0 or 1 in EDI/SEDI"
  thresholds_optimised_on: "training/validation only (never held-out test)"

rare_event_control:
  store_per_metric: [n_samples, n_events, n_nonevents]
  min_events: 10
  min_nonevents: 10
  sensitivity: [5, 10, 20]
  never_fill_masked_with_zero: true

event_matching:
  detect_on: "continuous daily sequences, per grid cell"
  match: ">=1 day overlap, one-to-one, maximise overlap / temporal IoU"
  unmatched_forecast: false_alarm
  unmatched_observed: miss
  duration_severity: "complete (non-censored) events only"

uncertainty:
  bootstrap: "year-block (or initialization-block) ONLY; never resample grid-days"
  ci: 0.95
  field_significance: "apply FDR / field significance where warranted"
  caution: "grid cells are a descriptive spatial distribution, not independent replicates"

regions:
  source_mask: {file: /work/projects/G09/QG.25.196/OBS/mask7reg_VnGAD_0.25.nc, var: vn_7regions}
  region_field: data/processed/region_codes_vngad.nc   # -1 outside, 0 UNASSIGNED, 1-7 R1-R7
  codes: [1, 2, 3, 4, 5, 6, 7]
  unassigned_code: 0            # D7: valid Tx cells with no R1-R7 code (n=45)
  national_domain: "all valid Tx cells INCLUDING region 0"     # maps, boxplots, national skill
  regional_R1_R7: "EXCLUDE region 0 until assignment resolved"  # region tables, lead curves
  diagnostics: [metadata/unassigned_region_cells.csv, outputs/figures/unassigned_region_cells.png]

leakage_rule:
  category_A_fixed: [P90, EHD_THRESHOLD]     # defined once 2003-2022; not leakage
  category_B_train_only: >
    predictor climatology, lead-dependent anomalies, means/std, normalisation,
    bias correction, calibration, feature selection, PCA/reductions, all ML and
    foundation-model params, early stopping, any optimised probability threshold.
    Held-out year enters NONE of these, including pretraining.
  transductive_all_years: "labelled side experiment only; never reported as out-of-sample skill"