Skip to contents

Validation and contributing

Source: vignettes/articles/validation.Rmd
validation.Rmd

What is checked

The package’s local validation compares the public implementation against the preserved reference engine on small scaled, correlated, high-dimensional, and budget-failure examples. Checks cover coefficient paths, fold losses, selected penalties, support counts, both starting points, preprocessing, numerical status, and original-scale predictions. Plot checks cover stored-fit displays and gaps for unavailable fits.

The latest local validation for version 0.1.10 passed 142 targeted checks and R CMD check with status OK. These are development checks, not CRAN acceptance or a guarantee of a global minimum. The documentation examples are also executed during the local site build.

Interpret numerical status

The objective is nonconvex. A returned fit is a stationary solution that meets the package’s numerical acceptance rules. The solver considers warm and zero starts, checks normalized stationarity, and reports active-curvature diagnostics. These checks do not certify global optimality.

  • ok describes the selected fit’s availability and numerical acceptance.
  • partial_search records unavailable CV candidates.
  • full_path_complete describes the stored full-training path.
  • cv_diagnostics and path diagnostics provide candidate-level details.

An unavailable candidate is ineligible for CV selection. Predictions and coefficient extraction reject an unverified selected fit. Plotting presents unavailable path positions as gaps.

Scope of the interface

The package supports dense finite matrices and Gaussian responses. It does not currently support observation weights, automatic missing-value handling, or a high-dimensional oracle guarantee. The shape is fixed and early stopping is not applied. Consult the function reference for defaults and controls.

Report a problem

Open an issue at the source repository. Include a minimal synthetic or public-data example, seed and folds when relevant, package version, sessionInfo(), controls, warnings, and fit status. For plots, include the device size. Do not attach confidential data.

The repository’s contribution guide describes local checks and the review of methodological or numerical changes.