NEWS

---

fdb 0.2.2

P3 now applies MCP to delta / se_delta, with lambda and gamma defined on that standardized scale. The smooth penalty integrates its slope to (sqrt(delta^2 + eps^2) - eps) / se_delta; eps remains in log-HR units. The transition half-width is rho_mcp * gamma_mcp * lambda in SE units. Both derivative chain factors are included in the raw curvature; negative curvature is still clipped only for the sandwich bread. The optimizer includes transition points and refines all detected local minima for P3.

This is a substantive replacement: recalibrate and reevaluate P3. Previous P3 lambdas, estimates, and result files do not describe this method. Other penalties and the fixed independent censoring design are unchanged. The returned lambda_eff is retained as the origin slope lambda / se_delta; it is no longer the lambda argument supplied to the MCP primitive.

fdb 0.2.1

• Censoring now uses a fixed exponential rate calculated from the internal population distribution, independent of realized replicate event times. Added calibrate_censor_rate() and optional censor_rate. Target zero yields exactly no censoring. Removed the stochastic per-replicate rate search.
• Simulation and calibration pass the resolved fixed rate to every replicate. Existing simulation results and calibrated lambdas must not be mixed with results from this revised generator.
• Fit and simulation method labels now match the manuscript. Short calibration API keys Li/P1/P2/P3/P4 remain supported; method_labels() translates keys and legacy output labels for presentation.

fdb 0.2.0 (2026-10-04)

• Prepare source for CRAN checks and regenerate help from roxygen comments.
• Add optional keep_raw retention to curve and study workflows.
• Include valid/missing counts and Monte Carlo standard errors in simulation summaries; keep rows for methods with entirely missing inference results.
• Stop the study wrapper when calibration fails instead of substituting 0.20.
• Honor explicit drift_set_confirm grids. NULL defaults to the calibration grid. Existing calls with identical calibration/confirmation grids retain their calibration target; calls previously passing a different ignored confirmation grid now use that grid and can produce different tuning.
• Retain coarse candidates in the fine calibration search.
• Default parallel runs to two workers, propagate installed library paths, clean up failed cluster initialization, and support NULL seeds in wrappers.
• Document conditional model-based SEs, approximate sandwich inference, finite-grid calibration, residual undercoverage and signed ESS gains.
• Penalty formulas and fixed-lambda fitting algorithms are unchanged by this release-preparation pass. Earlier manuscript changes to P2/P3 still require recomputation of results from the original penalty definitions.

Methodology updates

• Replace P2 drift-times-gate with the manuscript integrated-gate penalty.
• Smooth the MCP origin and flat-tail transition; expose rho_mcp (default 0.1) through fitting, calibration and study workflows.
• Match raw analytic curvature to the fitted penalties; retain clipping only in the plug-in variance calculation.
• Support no-covariate Cox fits and simulation (p = 0).
• Old P2/P3 fits and calibration results require recomputation.

fdb 0.1.0

• Initial release.
• User-facing functions: simulate_hybrid_cox, fit_internal_only, fit_naive_pooled, fit_li_adaptive_lasso, fit_P1_precision_L1, fit_P2_gated_L1, fit_P3_info_MCP, fit_P4_LRweighted_L1, fit_all_methods, fit_one_penalized_method, compute_sandwich_se, run_simulation, run_drift_curve, compute_ess_from_raw, add_ess_to_simulation_result, make_drift_set, make_drift_set_from_values, calibrate_lambda_grid, calibrate_lambda_grid_two_stage, calibrate_all_lambdas, and run_fdb_study.
• Reference scenario scenario_S1 and default tuning list lambdas_default exported for examples and tests.