DTEAssurance

• make_rpact_design_from_GSD_model(): for "PP" and "MatchedZ" futility, the efficacy critical values are taken from the efficacy-only design. A zero-alpha futility look cannot change them, but rpact's numerical integration moved them by about 8e-5 when the futility look was close to an efficacy look (e.g. 0.7 and 0.75).
• update_priors(): converged is now computed over finite R-hat values only and is NA (not TRUE) when none are finite.
• summarize_grid_by_kappa() and select_kappa_star() are replaced. power_LCB is now a one-sided 95% exact (Clopper-Pearson) lower bound (argument lcb_level, replacing conf_level); new columns n, n_NA_dropped, power_SE, ESS_SE; replicates with NA PP_val are dropped with a warning; exclude_nonconverged option. select_kappa_star() accepts per-scenario floors, treats an NA lower bound as infeasible, breaks ties in null ESS towards the larger kappa, and reports power and floors in the feasibility table.

New features

• single_paired_rep() and run_paired_scenario(): one simulated trial per replicate with one PP evaluation, recording interim, efficacy-look and final statistics (log-rank and modestly weighted log-rank) so that every design (D1-D5) and kappa is a post hoc summary. run_paired_scenario() runs in parallel chunks with an RDS checkpoint and resumes from it.
• compute_relative_floors(): D2 power minus a margin, per alternative scenario, for use as power_floor.
• update_priors(jags_seed = ) makes the MCMC reproducible, and new attributes rhat_max, n_nonfinite_rhat and n_retained_total.
• survival_test(return_HR = ): FALSE skips the Cox fit (used by the PP and group-sequential code, where the hazard ratio is not needed).
• calc_dte_assurance_adaptive() returns PP_val, P_Z1, P_Z2, P_Z3 per trial and a settings attribute (arguments, package and R versions, git commit if available, timestamp).
• Internal simulate_trial_with_recruitment(force_state = ) fixes the latent state, and the simulated data carry a truth attribute.
• inst/benchmark/time_one_rep.R (in the source repository only) times one paired replicate at the manuscript settings.
• make_gsd_design() (exported; previously the internal make_rpact_design_from_GSD_model(), kept as an alias for one release) and apply_GSD_to_trial() are now exported and documented.
• apply_design_rule(): D1-D5 decisions post hoc from run_paired_scenario() output, using the same decision labels as apply_GSD_to_trial(). A test checks that both give identical decisions, stop times and sample sizes on the same simulated trials.
• run_paired_scenario() is the single entry point for the paper's simulation tables.
• settings returned by the simulation functions now record all arguments plus the package, R, rjags and JAGS versions, the git commit of the working directory, the hostname and a timestamp.

Internal tidy-up (no change in results unless stated)

• One test wrapper (run_test()) and one truth simulator (simulate_trial_from_truth()) replace repeated code; outputs are identical under a fixed seed.
• Weibull control parameters from two landmark survival probabilities are computed in closed form (as in the update_priors() prior) instead of with nleqslv, which could fail silently. Results agree to about 1e-8 relative; invalid landmarks (S(t2) <= 0 or S(t2) >= S(t1)) now give an informative error. nleqslv moves from Imports to Suggests. The assurance Shiny app uses the same closed form.
• survival_test(): alpha has no default (it previously defaulted to 0.05; alpha = NULL returns Signif = NA when only Z is needed), the documented default alternative is now the actual default ("one.sided"), and Signif is always logical (FALSE if Z is NA).

History of earlier fixes (now removed from the code comments)

• The PP calculation evaluates the full group-sequential success event (crossing any remaining efficacy boundary, or success at the final analysis), not a single final test at a flat alpha.
• The trial evaluation in apply_GSD_to_trial() and calc_dte_assurance_adaptive() uses the specified test statistic (analysis_model$method) and the design's boundaries, replacing a hard-coded Cox Wald statistic at a flat qnorm(0.975); Success is derived only from Decision.
• "MatchedZ" futility looks are included in the design's information-rate grid with zero alpha spent.
• update_priors() monitors the latent state Z and returns R-hat and state-probability diagnostics; calibration output records them.
• The MW statistic from nphRCT::wlrt() is negated so that positive Z means benefit for all three methods.
• calibrate_PP_timing() defaults for update_priors_sims and PP_sims are 1000 and 2000 (previously hard-coded to 100 and 50).

Deprecated (kept, to be removed in a future release)

• The exponential control arm in PP_func() and update_priors(), the legacy censoring_model fallback of PP_func(), calibrate_PP_threshold() (and its internal single_calibration_rep()), and the internal summarize_gsd_results() and summarize_convergence().

Stricter inputs

• PP_func(): n_sims is required.
• calc_dte_assurance_adaptive(): update_priors_sims and PP_sims are required when futility_type = "PP".
• The PP code paths stop unless recruitment is uniform (recruitment_model$method = "power", power = 1), the assumption under which PP_func() draws future recruitment times.

Documentation

• update_priors(): n_samples is per chain (total n.chains * n_samples).
• "MatchedZ" futility is applied as binding (the trial stops).
• GSD_model$alpha_spending is user-specified cumulative alpha (typeOfDesign = "asUser").

Renamed to match the manuscript notation (no behaviour change)

• The predictive probability is now called PP and its futility threshold kappa, as in the manuscript:
• BPP_func() -> PP_func(); its return element BPP_df -> PP_df.
• calibrate_BPP_threshold() -> calibrate_PP_threshold() (returns PP_vec) and calibrate_BPP_timing() -> calibrate_PP_timing() (returns PP_values).
• summarize_grid_by_lambda() -> summarize_grid_by_kappa() (argument kappa_grid, column kappa) and select_lambda_star() -> select_kappa_star() (returns kappa_star).
• GSD_model$futility_type = "BPP" -> "PP" and GSD_model$BPP_threshold -> GSD_model$kappa.
• n_BPP_sims -> PP_sims in calc_dte_assurance_adaptive(), so the argument has the same name everywhere.
• The BPP_val column of the calibration output is now PP_val.
• Internally, the treatment-arm post-delay rate is now named lambda_e (in PP_func() and in the update_priors() JAGS models); the models are otherwise unchanged.
• A "Notation" section in the README maps manuscript symbols to package identifiers.

Deprecated (to be removed in the next release)

• BPP_func(), calibrate_BPP_threshold(), calibrate_BPP_timing(), summarize_grid_by_lambda() and select_lambda_star() still work but warn and forward to their replacements. Returned element and column names are not translated back to the old names.
• futility_type = "BPP" and GSD_model$BPP_threshold are accepted by calc_dte_assurance_adaptive() with a warning and mapped to "PP" and GSD_model$kappa.