The package examples evaluate the mathematical shape, project small deterministic regions, and reproduce one clinical scalar interval from aggregate summaries. They do not rerun the manuscript’s completed Monte Carlo study or recompute its benchmark methods. The completed study used 93 design cells and 5,000 replications per cell, for 465,000 datasets. Porting the numerical core does not replace those completed research outputs with new results.
With declared dependencies installed and a C++17 compiler available:
These commands build the package, its small vignettes, and its tests. The source package does not launch production simulations, bootstrap studies, benchmark sweeps, or model training.
The local port was compared with the preserved R research engine on 32 deterministic end-to-end cases, covering both calibrations, three- and six-dose inputs, flat and immediate-saturation boundaries, narrow intervals, empty intersections, exhausted budgets, and unresolved feasibility. The clinical interval was also compared with the saved manuscript result. This is a targeted equivalence gate, not exhaustive validation of all floating-point inputs or operating systems.
VALIDATION.md states the measured agreement and tested
platform. The historical R implementation remains the reference for the
completed study. Package tests additionally verify analytic endpoints,
numerical quadrature, F calibration, input validation, and
baseline/dose-unit invariance.
Keep the original observations, input units, variance definition,
calibration, alpha, tolerance, split budget, package version,
sessionInfo(), result status, endpoint gaps, and result
object. Save to a new location rather than overwriting an earlier
analysis. If stochastic sampling is part of a separate study, freeze its
seed and record that design separately: the projection functions
themselves do not draw random numbers.
The assumptions and finite-budget diagnostics remain essential. Neither the package nor its test suite proves coverage under heteroscedastic or dependent errors. Ordinary floating-point cushions do not establish a directed-rounding machine certificate. No speed advantage is claimed without a measured study.
The repository reproduction folder contains the original R projection functions, completed summary tables and a small clinical comparison script. It does not launch the production study.