Optimization Engineering: A Traceable Research-Software Workflow for Reproducible Numerical Optimization

Reproducible numerical optimization requires more than a solver termination flag: mathematical assumptions, evaluation budgets, software provenance, feasibility checks, and operational authority must remain distinguishable. We describe Optimization Engineering, an open-source collection of engineering workflows, documentation, verification scripts, and reference experiments. At the inspected repository revision, the project contains fourteen source-referenced agent-workflow specifications, eighty-four research prompts, numerical smoke tests, a gradient audit, a fixed-budget Rosenbrock benchmark driver, a figure generator, and continuous-integration configuration. We formalize the project’s evidence-gated workflow and demonstrate its executable component through a locally reproduced comparison of SciPy BFGS, Powell, and Nelder–Mead on a two-dimensional Rosenbrock function, each limited to 450 objective evaluations. The observed best objective values were 2.00559 × 10 −11 (123 evaluations), 1.20802 × 10 −1 (450 evaluations), and 1.09909 × 10 −18 (219 evaluations), respectively. These results corroborate one software smoke test; they do not establish statistical significance, universal algorithm rankings, implemented hardware safety, or cross-platform numerical identity. We distinguish implemented controls from proposed architectural safeguards and specify a validation path for future comparative and field studies