Changelog

Source: NEWS.md

BORG 0.3.2 (development)

Correctness fixes

• borg_inspect() now detects preprocessing leakage in base scale() output: a matrix carrying the "scaled:center"/"scaled:scale" attributes, or a data frame whose columns were replaced by scale() output. The stored centre and scale are compared with the training-row statistics, recovered from the scaled values themselves or from data. Standardising on all rows before the split is reported as a hard violation; standardising on the training rows only is not. borg_assimilate() re-standardises such a matrix from its unscaled values.
• borg_cv(strategy = "group_fold") now assigns groups to folds by their label rather than by position. Numeric group identifiers whose values were not 1:n_groups (e.g. site codes) previously produced empty test folds and full-data training sets.
• Stochastic fold generators (spatial_block, environmental_block, leave_location_out, llto, knndm, custom_block, random) now accept a seed argument threaded from borg_cv(seed = ) and borg_repeated_cv() instead of a hardcoded internal seed. Repeated CV now produces genuinely distinct fold sets across repeats for every strategy.
• borg_block_size() now varies the folds with the candidate block size (the block size drives the number of spatial clusters), so the reported optimum reflects the data rather than the search grid.
• Temporal block CV no longer backfills small early folds with future observations. Training is strictly past-only; the first block (with no prior data) is dropped, so v test blocks yield up to v - 1 usable folds.
• borg_conformal(method = "split") refits the model on the complement of the calibration set so the nonconformity scores are genuinely out-of-sample, and uses the conformal order statistic (type = 1) for the quantile. Split intervals now reach their nominal coverage.
• borg_di() / borg_aoa() follow Meyer & Pebesma (2021): the dissimilarity index is normalized by the mean pairwise training distance, and the applicability threshold is the outlier-adjusted maximum (Q75 + 1.5 * IQR) of the cross-validated training DI.
• Spatial buffering and distance-based diagnostics use Haversine distance when a geographic CRS is available. borg_cv() now threads the CRS to the buffer, and borg_block_size() / borg_check_residuals() accept a crs argument.
• borg_willmott() computes the refined index d1 with the correct denominator; borg_local_moran() uses conditional permutation for its pseudo p-values.
• The C++ duplicate-row check compares rows at full double precision instead of truncating to six decimals, and the compiled leakage checks validate their index arguments before dereferencing.

Internal

• Added parameter-recovery and coverage tests for the statistical estimators (borg_conformal, borg_aoa, borg_willmott, borg_calibration, borg_bootstrap, borg_debias).
• Consolidated the two metric dispatchers into a single implementation and the train-scope leak check into a shared helper; removed dead code.
• plot() and autoplot() for borg_result now offer the same set of plot types (split, risk, spatial, temporal, groups).

BORG 0.3.1

CRAN release: 2026-03-29

Nearest-neighbour distance matching (NNDM)

• New borg_cv(strategy = "nndm", prediction_points = ...) implements the leave-one-out NNDM scheme of Mila et al. (2022): each held-out point grows an exclusion radius until the test-to-train nearest-neighbour distance distribution matches the prediction-to-sample distribution, retaining at least half the data per fold. Complements the existing "knndm" k-fold form.
• Prediction-coordinate handling for "knndm" and "nndm" now shares one internal helper.

Clustered spatial sampling

• New borg_clustered_sample() draws parent-offspring clustered point samples over a bounding box, coordinate frame, sf polygon, or SpatVector. Useful for simulations and for reproducing the spatial dependence that random CV understates.

Normalized dissimilarity index

• borg_aoa(normalize = TRUE) adds a di_norm column rescaling the dissimilarity index by the AOA threshold, so di_norm = 1 marks the applicability boundary and values are comparable across models.

BORG 0.2.7

Suggested fixes in risk reports

• show(BorgRisk) now prints actionable suggested fixes for each detected risk
• as.data.frame(BorgRisk) includes a suggested_fix column
• Risks fixable by borg_assimilate() are labeled with the auto-fix call
• Manual-intervention risks provide specific guidance (e.g., “use spatial block CV”)

Improved ggplot2 autoplot methods

• autoplot(BorgRisk): Lollipop chart with fix annotations, point size encodes affected indices
• autoplot(borg_result, type = "temporal"): New temporal split visualization with look-ahead violation detection
• autoplot(borg_result, type = "groups"): New group assignment bar chart with overlap highlighting
• autoplot(BorgDiagnosis): Richer panel with detection threshold gauge bars

Better terra/sf spatial integration

• Spatial autoplot now uses tidyterra::geom_spatvector() for native SpatVector rendering when available
• sf objects preserve geometry directly (no coordinate extraction + reconstruction round-trip)
• Faceted spatial CV plots work natively with both sf and SpatVector inputs
• Added tidyterra to Suggests

---

BORG 0.2.3

Native R Integration

This release replaces custom functions with standard R S3 methods for a more idiomatic interface.

S3 methods for plotting

• plot(BorgRisk): Visualize risk assessment results
• plot(borg_result): Visualize CV fold splits
• plot(borg_comparison): Compare random vs blocked CV results

S3 methods for summaries

• summary(BorgDiagnosis): Generate methods section text for publications
• summary(BorgRisk): Summarize detected risks
• summary(borg_result): Generate methods text from borg() output
• summary(borg_comparison): Detailed comparison summary

Dedicated export functions

• borg_certificate(): Create structured validation certificate
• borg_export(): Write certificate to YAML/JSON file

Removed

• borg_plot() - use plot() instead
• borg_report() - use summary(), borg_certificate(), or borg_export()
• Legacy plot functions (plot_split, plot_risk, etc.)
• Legacy report functions (borg_methods_text)

---

BORG 0.2.2

New features

Framework wrappers

BORG-guarded versions of common CV functions that block random CV when dependencies detected:

• borg_vfold_cv(): Wraps rsample::vfold_cv() with dependency checking
  • Blocks random CV when spatial/temporal/clustered dependencies detected
  • auto_block = TRUE automatically switches to appropriate blocked CV
  • allow_override = TRUE allows proceeding with warning
• borg_group_vfold_cv(): Wraps rsample::group_vfold_cv() with additional checks
  • Warns about dependencies not handled by grouping alone
• borg_initial_split(): Wraps rsample::initial_split()
  • Enforces chronological splits when time specified
  • Warns about spatial dependencies
• borg_trainControl(): Wraps caret::trainControl()
  • Blocks random resampling methods with detected dependencies

Hook system (experimental)

• borg_register_hooks(): Register validation hooks on framework functions
• borg_unregister_hooks(): Remove registered hooks

---

BORG 0.2.1

New features

Empirical CV comparison

• borg_compare_cv(): Run random vs blocked CV on the same data to empirically demonstrate metric inflation
  • Supports spatial, temporal, and clustered data
  • Custom model and predict functions
  • Multiple metrics: RMSE, MAE, R², AUC, accuracy
  • Paired t-test for statistical significance
  • Plot methods: boxplot, density, paired comparison

Publication-ready reporting

• borg_methods_text(): Generate copy-paste methods section text for manuscripts
  • Includes dependency diagnosis details
  • Cites quantitative metrics (Moran’s I, ICC, etc.)
  • Optional empirical comparison results
  • BORG package citation
• borg_certificate(): Create structured validation certificates
  • Machine-readable format
  • Includes data characteristics, diagnosis, and inflation estimates
  • Timestamps and version tracking
• borg_export(): Export certificates to YAML or JSON format

API improvements

• Unified borg() entry point with two modes:

  • Diagnosis mode: borg(data, coords=, time=, groups=) returns diagnosis + CV folds

  • Validation mode: borg(data, train_idx=, test_idx=) validates existing splits

• Renamed parameters for consistency: temporal_col → time, group_col → groups, spatial_cols → coords

---

BORG 0.2.0

Major changes: From Validator to Enforcer

BORG now detects data dependency structure and enforces appropriate cross-validation strategies. Random CV is blocked when dependencies are detected.

Dependency diagnosis

• borg_diagnose(): Automatically detects data dependency structure
  • Spatial autocorrelation: Moran’s I test, variogram-based range estimation
  • Temporal autocorrelation: ACF/Ljung-Box test, decorrelation lag detection
  • Clustered structure: ICC computation, design effect estimation
  • Returns severity assessment and recommended CV strategy
  • Estimates metric inflation from using random CV
• BorgDiagnosis S4 class: Structured diagnosis results with slots for:
  • dependency_type: “none”, “spatial”, “temporal”, “clustered”, “mixed”
  • severity: “none”, “moderate”, “severe”
  • recommended_cv: appropriate CV strategy
  • inflation_estimate: estimated AUC/RMSE bias from random CV

Enforced CV generation

• borg_cv(): Generates valid CV schemes based on diagnosis

  • Spatial blocking: k-means clustering with block size > autocorrelation range

  • Temporal blocking: chronological splits with embargo periods

  • Group CV: leave-group-out with balanced fold assignment

  • Mixed strategies: spatial-temporal, spatial-group, temporal-group

  • Random CV disabled when dependencies detected (requires explicit allow_random = TRUE)

  • Output formats: list, rsample, caret, mlr3

New features

Visualization functions

• plot_split(): Visualize train/test split distribution with temporal or group structure
• plot_risk(): Display risk assessment results as horizontal bar chart
• plot_temporal(): Timeline visualization with gap analysis and look-ahead detection
• plot_spatial(): Spatial split visualization with convex hulls
• plot_groups(): Group-based split visualization with leakage highlighting

Model inspection

• Extended borg_inspect() to support fitted model objects:
  • lm and glm models (checks data used in fitting)
  • ranger random forest models
  • xgboost models
  • lightgbm models
  • parsnip model fits
  • workflow objects (tidymodels)

Audit functions

• audit_predictions(): Validate prediction vectors against expected indices
• cv_leakage_report(): Generate detailed cross-validation leakage reports
• audit_importance(): Detect feature importance computed on test data (SHAP, permutation)

Tidymodels integration

• Added tune_results inspection for tidymodels tuning objects
• Detects when hyperparameter tuning uses test data in resamples

Configuration

• Added borg_auto_check() to enable/disable automatic validation
• Added borg_options() to query current configuration
• New options: borg.auto_check, borg.strict, borg.verbose

Bug fixes

• Fixed rsample vfold_cv inspection for attribute structure changes

BORG 0.1.0

Initial release.

Core functionality

• borg_guard(): Creates a validation context for train/test splits with support for temporal, spatial, and grouped structures
• borg_validate(): Comprehensive workflow validation including:
  • Index overlap detection
  • Duplicate row detection
  • Target leakage detection (suspicious naming, perfect separation)
  • Feature engineering leakage (global standardization, rank features)
  • Threshold selection leakage
  • Spatial autocorrelation checks
  • HPO validation checks (test data usage, non-nested CV)
• borg_inspect(): Inspects preprocessing objects for data leakage:

  • caret preProcess objects

  • caret trainControl objects

  • tidymodels recipe objects

  • prcomp PCA objects

  • rsample resampling objects

• borg_assimilate(): Assimilates leaky pipelines into compliance (auto-fix)
• BorgRisk S4 class for structured risk assessment reports

Performance

• C++ backends via Rcpp for fast hash-based index overlap and duplicate detection

Framework integration

• caret: preProcess, trainControl, train objects
• tidymodels: recipe, rsplit, vfold_cv, rset objects
• mlr3: task and resampling validation
• Base R: manual index-based splits