Theory-Constrained Additive Modeling with Finite-Sample Uncertainty and Paired Risk Auditing

This archive accompanies the technical report “Theory-Constrained Additive Modeling with Finite-Sample Uncertainty and Paired Risk Auditing.” The project develops a structured regression framework integrating three components: Theory-constrained additive modeling with monotonic shape restrictions, Finite-sample predictive uncertainty via split conformal calibration, Paired risk auditing against unconstrained baselines under identical data splits. Rather than focusing exclusively on predictive accuracy, the framework emphasizes structural discipline, calibration validity, and transparent evaluation of bias–variance tradeoffs. Shape constraints are imposed according to domain-informed expectations and evaluated for stability across repeated train/test splits. Predictive intervals are constructed using distribution-free conformal methods and empirically validated for coverage. Performance tradeoffs introduced by structural constraints are assessed through paired ΔRMSE comparisons computed on identical resampling splits. The archive contains the full technical report, cleaned figures, performance summaries, paired risk audit outputs, and serialized model objects sufficient to verify all reported quantitative results. The original exploratory scripts are not included; however, all numerical results in the report correspond directly to the archived artifacts provided. This work is intended as a methodological technical report emphasizing reproducibility, interpretability, and controlled performance comparison.