Changes in version 0.1.2 (2026-06-24)                  

CRAN resubmission — no code changes; fresh win-builder check to clear a
stale build-timestamp note.

                        Changes in version 0.1.1                        

CRAN resubmission addressing reviewer feedback.

Documentation

  - Replaced \dontrun{} with \donttest{} in the brm() example and
    unwrapped the learner() example (it runs instantly).

Internal changes

  - Removed hardcoded set.seed() calls inside choose_num_blocks() and
    best_kmeans(); these were unnecessary and modified the caller's RNG
    state.
  - simulate_blockwise_missing() no longer calls set.seed() directly.
    The seed argument now defaults to NULL (use the caller's RNG); when
    supplied, the seed is applied locally via withr::with_seed() so the
    caller's RNG state is preserved.
  - Added withr to Imports.

                        Changes in version 0.1.0                        

First public release. Initial CRAN submission.

Features

  - brm() — fit a Blockwise Reduced Modeling ensemble (S3 class "brm").
  - predict.brm() — route test instances to their best-matching subset
    model.
  - choose_num_blocks() — elbow heuristic for the number of blocks.
  - learner() — learner-agnostic fit/predict specification; convenience
    builders for linear models (learner_lm, learner_glm_binomial), trees
    (learner_rpart), random forests (learner_ranger), and gradient
    boosting (learner_gbm).
  - simulate_blockwise_missing() — mask complete data with a blockwise
    missing pattern for benchmarking.
  - Bundled datasets: bike, adult, house — the three benchmark datasets
    used in Srinivasan, Currim, and Ram (2025)
    doi:10.1287/ijds.2022.9016.