delta_lsi(): inference tier strings renamed to
accurately reflect what each tier provides. "C_point_only"
→ "C_signflip" (the sign-flip p-value is available at this
tier, not just point estimates); "B_ci_only" →
"B_signflip_ci" (both the sign-flip p-value and BCa CI are
available). Code that compares result@tier against the old
string literals must be updated.delta_lsi() gains a block_size argument
and makes exchangeability actionable for
"blocked_time" inputs. When
exchangeability = "blocked_time", the sign-flip test now
uses a block procedure that flips contiguous blocks of repeats together,
preserving serial autocorrelation under the null.
block_size is auto-estimated from the AR(1) of the
repeat-level deltas when NULL (default) and capped at
floor(R/3) to guarantee at least three independent blocks.
The @info slot gains block_size_used and
n_blocks fields. If the block structure yields fewer than
five independent blocks, @p_value is set to NA
and a warning is issued.delta_lsi() now emits an explicit warning when
exchangeability is "by_group" or
"within_batch", informing users that those modes are stored
but inference still uses the iid sign-flip procedure. Previously these
values were accepted silently without affecting computation.fit_resample(): compact + combined mode now correctly
excludes constraint-axis violations from training sets. Previously the
compact fallback used setdiff(all, test), ignoring
multi-axis constraints declared via
make_split_plan(constraints = ...). The same fix is applied
in the as_rsample() conversion path for consistency.delta_lsi(): R_eff and the inference tier
are now recomputed after repeat-level intersection, so that dropped
all-NA repeats correctly reduce the effective sample size and select the
appropriate tier.fit_resample(): fold error messages are now correctly
captured when running in parallel via future.apply.
Previously <<- mutations inside worker processes were
silently lost; errors are now attached as result attributes and
extracted after the parallel map.tune_resample(): fold-ID columns (id,
id2, .notes) no longer leak into
hyperparameter aggregation in the internal select_config()
helper.summary.LeakFit() now returns
object@metric_summary invisibly, matching the documented
return value (previously returned the object itself).bioLeak-intro) referencing a shadowed
data frame for sample count; now reads from
fit_safe@splits@info$coldata.audit_leakage() roxygen documenting a
duplicates column named in_train_test; the
actual column name is cross_fold.make_split_plan(): time-series mode now warns and skips
folds with fewer than 3 test samples instead of producing degenerate
folds.fit_resample(): added bounds checking for
repeat_id in compact fold resolution to produce a clear
error instead of a cryptic index failure.show() and summary() for
LeakDeltaLSI now label the sign-flip p-value as testing
mean(Δr) (delta_metric), not delta_lsi, making the
estimator–inference pairing explicit.summary() prints a diagnostic note when the sign-flip
p-value and BCa CI lead to qualitatively different conclusions (one
significant, one spanning zero), which can occur when outlier repeats
pull the arithmetic mean away from the Huber estimate.summary() prints the block size and number of blocks
used when exchangeability = "blocked_time".constraints in make_split_plan(), generalizing
beyond two-axis combined CV while preserving train/test exclusion across
all declared axes.compact = TRUE split storage (fold assignments)
for large datasets to reduce split object memory footprint.check_split_overlap() for explicit
overlap-invariant validation across fold/group axes.cv_ci() (with Nadeau-Bengio correction) and
integrated CI columns into fit_resample() and
tune_resample() metric summaries (*_ci_lo,
*_ci_hi).guard_to_recipe() to map guarded preprocessing
configurations to recipes pipelines with explicit
fallback/warning behavior.benchmark_leakage_suite() for reproducible
modality-by-mechanism benchmark grids and detection-rate summaries.audit_leakage() diagnostics with mechanism
taxonomy fields (mechanism_class, taxonomy,
mechanism_summary) and richer risk attribution
outputs.p_value_adj,
flag_fdr) with selectable multiple-testing correction
(target_p_adjust, target_alpha).feature_space
(raw/rank) and duplicate_scope
(train_test/all) controls for duplicate
diagnostics.perm_mode handling for rsample-derived splits and safer
perm_refit = "auto" behavior.split_cols = "auto",
mode/perm-mode propagation, stricter compatibility checks).tune_resample(): final
refit now aggregates hyperparameters across outer folds
(median/majority) instead of selecting a single best outer fold.tune_resample() using inner-fold predictions
(tune_threshold, threshold_grid,
threshold_metric).fold_status) and
elapsed timing in both fitting and tuning paths for better failure-mode
observability.bioLeak.strict, bioLeak.validation_mode) with
structured condition classes for safer recipe and workflow
guardrails..bio_capture_provenance) and
attached provenance metadata to LeakFit,
LeakAudit, and LeakTune.summary.LeakAudit() output with explicit
Mechanism Risk Assessment reporting.fit_resample() to
avoid fold-time failures when recipes reference split metadata columns
(for example subject).simulate_leakage_suite() default B,
auto refit cap handling).paper/ with
refreshed large-scale simulation outputs and case-study artifacts.tune_resample(): nested cross-validation using tidymodels
tune/dials with leakage-aware outer
splits.fit_resample() now accepts rsample rset/rsplit
objects as splits, recipes::recipe for
preprocessing, workflows::workflow as learner,
and yardstick::metric_set for metrics.
as_rsample() converts LeakSplits to an
rsample rset.learner argument in fit_resample().calibration_summary() and plot_calibration()
for probability calibration checks;
confounder_sensitivity() and
plot_confounder_sensitivity() for sensitivity
analysis.simulate_leakage_suite() for generating controlled leakage
scenarios and benchmarking audit sensitivity.audit_report():
renders a self-contained HTML summary of all audit results for sharing
and review.audit_leakage_by_learner() to audit each learner in a
multi-model fit separately.audit_leakage() for supported tasks, complementing the
existing univariate scan.perm_refit = TRUE or "auto") in
audit_leakage() for a more powerful permutation gap test
when refit data are available.fit_resample() for imbalanced classification tasks.plot_fold_balance(),
plot_overlap_checks(),
plot_perm_distribution(),
plot_time_acf().LeakSplits, LeakFit,
LeakAudit) now include setValidity checks for
slot consistency.summary() methods for LeakFit,
LeakAudit, and LeakTune improved with clearer
console output and edge-case handling.impute_guarded() gains enhanced diagnostics and RNG
safety..guard_fit() and .guard_ensure_levels()
made more robust with better error messages.permute_labels) gains
verbose mode, digest-based caching, and improved stratification
safety.audit_leakage() handles NA metrics gracefully and
enriches trail metadata.make_split_plan() improved stratification logic and
reproducible seeding.audit_report() now renders from a temporary copy of the
Rmd template to avoid write failures on read-only file systems
(e.g. during R CMD check).bioLeak-intro) rewritten with
guided workflow and leaky-vs-correct comparisons.fit_resample() result aggregation when folds fail
during preprocessing.missForest preprocessing dropping rows.glmnet folds receiving non-numeric design
matrices.make_split_plan() for
leakage-aware splitting (subject-grouped, batch-blocked, study
leave-out, time-ordered); fit_resample() for
cross-validated fitting with built-in guarded preprocessing (train-only
imputation, normalisation, filtering, feature selection).audit_leakage() with
label-permutation gap test, batch/study association tests, univariate
target leakage scan, and near-duplicate detection.impute_guarded(), predict_guard(),
.guard_fit(), .guard_ensure_levels().LeakSplits, LeakFit,
LeakAudit.glm, glmnet,
ranger, xgboost (via
custom_learners).SummarizedExperiment input support.
Need a high-speed mirror for your open-source project?
Contact our mirror admin team at info@clientvps.com.
This archive is provided as a free public service to the community.
Proudly supported by infrastructure from VPSPulse , RxServers , BuyNumber , UnitVPS , OffshoreName and secure payment technology by ArionPay.