The RcppArmadillo implementation was validated locally on Apple arm64 with R 4.6.0, Apple clang 17, Rcpp 1.1.1.1.1, and RcppArmadillo 15.2.7.1.
Targeted equivalence gate
- 32 deterministic end-to-end comparisons against the preserved R engine.
- Both centered contrast and full mean-region calibration.
- Three- and six-dose stored validation inputs and hand-specified boundary, narrow, empty, budget-exhausted, and unresolved-feasibility cases.
- The maximum absolute difference between interval endpoints was 8.881784197001249e-16; the required comparison tolerance was 1e-9.
- Status, convergence, and termination agreed in every comparison.
- Analytic envelopes and fixed-shape slices agreed within the required tolerance.
- Shape, integrated response, derivatives, and quadratic hull primitives were checked against the R reference, including the closed endpoints.
- The saved 369-observation clinical interval was reproduced without changing the data or fitting any comparator.
- Projection calls preserved the random-number state; no random draws were used.
- Historical source and selected input/result seals were unchanged.
Package tests
Tests cover independent numerical quadrature, analytic model closures, the Gaussian F calibration, raw/summary consistency, missing-input rejection, corrupted-region rejection, target containment at boundaries, numerical status, and invariance to a baseline shift and a dose-unit change.
The small tests and documentation examples are not new Monte Carlo evidence. No production run or speed benchmark was performed. Windows and Linux compilation have not yet been tested. The enclosing inequalities are exact-arithmetic arguments evaluated in ordinary double precision; no formal directed-rounding certificate or independent external proof review is claimed.