Residual-First Benchmarks for Constrained-Versus-Baseline Regression Under Declared Standard Uncertainty: Reproducibility Kit (v1.0.0)

This repository contains the complete reproducibility kit accompanying the manuscript: J. D. Bourassa, “Residual-First Benchmarks for Constrained-Versus-Baseline Regression Under Declared Standard Uncertainty” The kit implements a fully specified, deterministic benchmark protocol for adjudicating constrained-versus-baseline regression models in measurement applications under declared, GUM-consistent standard uncertainty. The workflow operates entirely in observation space and regenerates: Weighted least-squares (WLS) fits for baseline and constrained models Normalized residual diagnostics ( z = r / u(y) ) Residual structure tests (tails, runs, trend diagnostics) Model-comparison metrics (AIC, BIC, nested-model F-test, k-fold cross-validation) Declared modeling sensitivity ( S(x) = |\hat{y}_b - \hat{y}_c| / u(y) ) Machine-readable benchmark records (JSON) documenting the full decision record Two worked examples are included: Digitized curve example (methodological only) — Demonstrates protocol behavior under declared digitization uncertainty. Heteroscedastic calibration example — Weighted regression with temperature-dependent standard uncertainty, representative of instrumentation calibration practice. The repository contains: Input datasets (CSV) A single entry-point execution script (src/main.py) Regenerated tables and figures Sensitivity profiles ( S(x) ) Formal benchmark records in JSON format Dependency specification (requirements.txt) All results are deterministic given the provided datasets and declared uncertainty models. This archive is versioned and DOI-minted to support transparent, auditable methodological use in measurement science.