Run one kpe2 shared-nuisance cross-fit pass.

Partition the rows into n_folds folds. Tile them into K/b non-overlapping evaluation blocks of size b = n_folds - n_nuisance_folds. For each block: fit ALL nuisances (reward model, contextual policy, best single arm, and – when is_rct = FALSE – the propensity p(a|x)) on the other m = n_nuisance_folds folds, then evaluate the AIPW influence on the block. Each sample is evaluated exactly once per shuffle. Default n_nuisance_folds = n_folds - 1 gives leave-one-fold-out.

Usage

.single_split_kpe(
  inputs,
  reward_builder,
  policy_builder,
  best_arm_builder,
  n_folds,
  seed,
  n_nuisance_folds = NULL,
  is_rct = TRUE,
  propensity_builder = NULL,
  folds_override = NULL
)

Arguments

folds_override

Optional list of length n_folds, each a vector of 1-based row indices. Used by the cross-language parity test to inject sklearn's KFold shuffle so both languages operate on identical folds.