NEWS

---

spfcICOMP 0.1.0.9000

• Made benchmark printing response- and validation-aware, so structurally inapplicable metrics are no longer displayed as missing values.
• Separated applicable covariance diagnostics from the main performance table while retaining the complete internal benchmark schema for backwards compatibility.
• Added regression tests for compact continuous, categorical, and cross-validation benchmark summaries.
• Retained CITATION.cff for GitHub and Zenodo while excluding it from R source-package builds to prevent non-standard-file check notes.
• Opened the post-release development cycle after publication of version 0.1.0 on CRAN.
• Added CRAN and Zenodo badges and made CRAN installation the primary README route.
• Recorded the version-specific Zenodo DOI (10.5281/zenodo.22310095) and stable concept DOI (10.5281/zenodo.22310094).
• Added machine-readable citation metadata for the archived 0.1.0 software release.

spfcICOMP 0.1.0 (2026-09-03)

• Began GitHub and CRAN release hardening without changing the statistical methodology or thesis-result implementation.
• Added repository metadata, README, citation metadata, and CRAN-preparation notes.
• Excluded thesis-analysis artefacts and exploratory outputs from the CRAN source tarball while retaining them for GitHub reproducibility.
• Replaced over-strong public wording about sparsity/selection with feature-screening terminology where appropriate.
• Removed third-party Riboflavin and Golub benchmark copies from the distributable package; the GitHub thesis-analysis layer now obtains Riboflavin from the archived CRAN hdi 0.1-10 source (with an installed-hdi fast path when available) and Golub from Bioconductor multtest when empirical reproduction is requested.
• Passed a local CRAN-style check with 0 errors, 0 warnings, and 0 notes and established a five-platform GitHub Actions R CMD check workflow.
• Hardened validation stratification so numeric binary responses are treated as categorical strata in train-test splitting and cross-validation fold construction.
• Added regression tests for numeric binary-response stratification without changing the statistical methodology or thesis-result implementation.
• Verified 149/149 local tests, a local R CMD check with 0 errors, 0 warnings, and 0 notes, exact R 3.5.0 compatibility, and the five-platform GitHub Actions matrix at commit fc80e696b7a64a2b8ab7c274f0af6c13e0fbb8ff.
• Made citation metadata safe to evaluate before package installation by removing the runtime package-description lookup.
• Declared British English and added a reviewed spelling wordlist for technical terms, acronyms, proper names, and method identifiers.
• Built and verified the exact 0.1.0 source tarball with 149/149 test expectations passing, no spelling flags, no missing or forbidden source-package entries, and a Windows R 4.6.0 R CMD check result of 0 errors, 0 warnings, and 0 notes.

spfcICOMP 0.0.0.9008

• Align the standalone validity-check subspace tolerance with the package unit-test tolerance (1e-7); no methodological or simulation logic changed.

• Relaxed the numerical tolerance in the DGP subspace-equivalence unit test from 1e-8 to 1e-7; no simulation or estimation logic changed.

spfcICOMP 0.0.0.9006

• Replaces the collapsed forward continuous simulation with a model-faithful inverse PFC DGP whose nominal d=2 inverse mean has rank two by construction.
• Constructs the inverse loading so span(Sigma^{-1} Gamma) equals the sparse true reduction subspace, preserving valid active-variable evaluation under correlated AR(1) residual covariance.
• Defines SNR in the true sufficient coordinates and scales the inverse mean so the requested SNR is exact in population and invariant to p.
• Generates an independent test sample for every continuous replicate and evaluates RMSE/MAE out of sample only.
• Uses a fixed degree-2 continuous response basis in thesis simulation runners, sufficient to represent both d=1 and d=2 without using the true d to tune the fitted basis.
• Adds regression tests guarding against rank collapse and accidental in-sample prediction evaluation.
• Explicitly restricts the current binary PFC simulation to d=1 rather than silently treating a rank-one binary response basis as a multi-dimensional recovery problem.

spfcICOMP 0.0.0.9005

• Froze the thesis C1F-to-L1 calibration as icomp_hd_floor.
• The new default global row penalty is median(row_norms) * sqrt(2*log(p)/n) * (1 + C1F/(1+C1F)).
• This separates the high-dimensional multiplicity/noise floor from the bounded covariance-complexity multiplier without introducing a user-tuned LASSO lambda.
• Retained the 0.0.0.9004 robust_universal and historical raw calibrations for reproducibility.
• Added package tests for the exact frozen penalty formula and default calibration.
• Added high-dimensional-factor and complexity-multiplier diagnostics to simulation summaries.

spfcICOMP 0.0.0.9004

• Recalibrated C1F-driven row-group sparsity with a scale-compatible, deterministic high-dimensional penalty.
• Removed the 1/p attenuation caused by sum-to-one adaptive-weight normalisation in the thesis selection route.
• Preserved the raw C1F route as a reproducibility option.
• Fixed simulation summaries so SNR and categorical signal-strength levels remain separate factorial cells.
• Added C1F penalty diagnostics to simulation output and smoke-test reporting.

spfcICOMP 0.0.0.9001

• Added genuine structural-dimension recovery to simulation design runs. The true structural dimension is no longer supplied to the estimator in the thesis workflow; AIC, BIC, CAIC, ICOMP(IFIM), and CICOMP each select d from a candidate grid.
• Added C1F-driven adaptive row-L1 feature selection via selection_rule = "c1f", while retaining legacy quantile thresholding for backwards compatibility.
• Added dimension recovery indicators, C1F penalty values, and confusion-count fields to simulation outputs.
• Updated simulation summaries to report dimension recovery rate and selected-d mean/SD by criterion.
• Added a thin thesis_analysis/ reproducibility workflow built entirely on package functions.