EbD RAG Design Problem

Simulation harness, analysis scripts and derived output reproducing every number, table and figure of a study that recasts the design of an ethical-by-design retrieval-augmented generation (RAG) pipeline as a constrained mixed-integer multiobjective program. A composite Trust Gap Score over privacy, opacity and bias sub-scores is minimized over four decision variables — the differential-privacy budget, the hybrid retrieval weight, the fair-ranking intervention strength, and a binary selectionover two cryptographic privacy-enhancing technologies — subject to retrieval-quality, latency, demographic-parity and privacy-budget constraints. The package solves the program by exhaustive enumeration over 2,039,184 configurations, traces the Pareto front by the epsilon-constraint method, verifies the solution with an independent NSGA-II control run (hypervolume agreement within0.95%), and decomposes the variance of four outputs over all sixteen stipulated model constants using Sobol indices with Saltelli sampling and bootstrap confidence intervals (18,432 complete program solutions per perturbation width). STATUS OF THE NUMBERS. Every constant in the model is stipulated rather than estimated from data. The results characterize the structure of the design problem — which constraints bind, how the levers interact, where the feasible set ends, and how far each conclusion depends on each stipulation — and do not describe the privacy, fairness or retrieval quality of any deployed system. The global sensitivity analysis is therefore a primary result of the package rather than arobustness appendix. CONTENTS src/ the model and five analysis stages run_all.py single-command reproduction driver results/ derived .npz arrays and .json summaries tables/ the manuscript tables as .csv, including full Sobol indices with bootstrap intervals for every parameter-output pair figures/ all figures as .pdf and .png paper/ manuscript source and compiled PDF submission/ journal submission set and an automated check of the manuscript against the target journal's author requirements REPRODUCING python -m pip install -r requirements.txt python run_all.py --clean A verification script (src/verify.py) independently re-derives all 118 numerical values asserted in the manuscript directly from the model, rather than reading them back from the saved result files, and reports any mismatch. Everything is deterministic given the fixed seed (42): repeated runs produce byte-identical output, figures included. The package requires no network access, no external data and no GPU, and reproduces in roughly three minutes on a recent multi-core machine. Reproduced with Python 3.13, NumPy 2.5, SciPy 1.18, pandas 3.0, Matplotlib 3.11 andpymoo 0.6.2.