Calibrated Regression Model with Conformal Uncertainty and Robustness Diagnostics

This archive contains a serialized regression model and associated evaluation artifacts implementing a calibrated prediction system with structured robustness diagnostics. The core object in this release is a trained additive regression model (model.rds) accompanied by structured performance outputs and uncertainty calibration diagnostics. The modeling framework integrates: Additive regression with controlled smoothness, Split conformal calibration for distribution-free predictive intervals, Empirical coverage validation on held-out evaluation data, Subgroup-level performance assessment, Permutation-based feature importance analysis, Gradient-based sensitivity diagnostics, Robustness curve evaluation under varying perturbation regimes. The conformal calibration layer computes nonconformity scores based on absolute residuals and constructs two-sided predictive intervals with nominal coverage level 1−α1 - \alpha1−α. For α = 0.1, empirical coverage on held-out data is approximately 0.923, demonstrating calibration behavior under finite samples. Overall predictive accuracy is summarized via RMSE (≈ 1.97 across 300 evaluation observations), with subgroup-level RMSE values provided for diagnostic comparison. Permutation importance outputs quantify marginal contribution of predictors to predictive error, while gradient sensitivity analysis evaluates local response behavior under feature perturbation. Robustness curves summarize stability of predictive performance across evaluation settings. The purpose of this release is reproducibility and transparency. It archives the trained model object, calibration outputs, subgroup metrics, and robustness diagnostics sufficient to verify reported evaluation statistics. The archive does not include a manuscript or full modeling pipeline; rather, it serves as a computational artifact release documenting model behavior, uncertainty calibration performance, and robustness characteristics. This artifact may be useful for researchers or practitioners interested in calibrated regression systems, finite-sample uncertainty evaluation, structured model diagnostics, and reproducible evaluation workflows.