NEWS

---

ml 1.0.0

First stable CRAN release. The Hastie workflow in R.

Core workflow

• Four-verb grammar: ml_split(), ml_fit(), ml_evaluate(), ml_assess(). The evaluate/assess boundary prevents data leakage by separating iterative model selection from final generalization estimates.
• Three-way data splitting (train/valid/test) with automatic stratification for classification targets.
• Mandatory reproducibility: seeds are auto-generated and stored when not provided explicitly.
• Per-fold preprocessing: encoding and scaling fit on training folds only.

Algorithms

• 13 algorithms: logistic regression, linear regression (ridge), decision tree, random forest, extra trees, gradient boosting, histogram gradient boosting, XGBoost, KNN, naive Bayes, SVM, elastic net, AdaBoost.
• Optional backends via Suggests: xgboost, ranger, glmnet, e1071, kknn, naivebayes, rpart. Without these, algorithms fall back to the optional Rust engine or report as unavailable.

Additional verbs

• ml_screen() for rapid algorithm comparison across all available backends.
• ml_tune() for hyperparameter tuning with random search and cross-validation.
• ml_stack() for model ensembling via out-of-fold stacking.
• ml_drift() for data drift detection (KS test).
• ml_shelf() for model staleness monitoring.
• ml_explain() for feature importance (impurity-based and coefficient-based).
• ml_calibrate() for probability calibration (Platt scaling).
• ml_validate() for pass/fail gating against user-defined rules.
• ml_profile() for dataset profiling.
• ml_save() / ml_load() for model persistence in .mlr format.

Rust backend

• Optional compiled Rust engine via extendr bindings. Detected at install time by configure; no Rust required for a fully functional installation.

ml 0.2.0

Changed

• ml_assess() terminal constraint now enforced per-partition via content-addressed fingerprinting. Serialization and deepcopy bypasses are closed.

New

• ml_cv(), ml_cv_temporal(), ml_cv_group() for cross-validation.
• ml_verify() for post-fit model verification.
• ml_prepare() for explicit preprocessing.
• Content-addressed provenance registry (rlang::hash fingerprinting).
• Cross-verb provenance checks (train/test from same split).

Fixed

• Per-holdout assess enforcement: ml_assess() now rejects a second call on the same test partition regardless of which model calls it.
• Fixed ml_prepare() return value extraction (X and norm fields).

ml 0.1.2 (2026-03-19)

New

• ml_cv(), ml_cv_temporal(), ml_cv_group() for cross-validation.
• ml_verify() for post-fit model verification.
• ml_prepare() for explicit preprocessing.
• Content-addressed provenance registry (rlang::hash fingerprinting).
• Cross-verb provenance checks (train/test from same split).

Fixed

• Per-holdout assess enforcement: ml_assess() now rejects a second call on the same test partition regardless of which model calls it. The provenance registry tracks spent holdouts via content-addressed fingerprinting.
• Fixed ml_prepare() return value extraction (X and norm fields).