TcScreening: leakage-aware, uncertainty-calibrated machine-learning screening of superconducting critical temperature - full results and code

Complete result set and code for the study "Composition-based screening of superconducting critical temperature: an interpretable, uncertainty-calibrated machine-learning framework evaluated under leakage-controlled validation" (Materials Futures, Focus Issue: Artificial Intelligence for Materials Science). Contents: 630 controlled fits (4 split regimes × 6 models × 5 seeds × 5 folds; RMSE/MAE/R2, split-conformal 90% coverage and widths), Holm-corrected paired significance tests, feature-set and top-k ablations, deduplication sensitivity, SHAP attributions and stability, stylized perturbation probes, residuals, 3,786,600 out-of-fold prediction rows, post-review analyses (family census, test-size-weighted means, Mondrian conformal, instance-distance diagnostic, normalization ablation), run manifests with package versions and scope SHA-256, paper figures (vector PDF + PNG), and a code snapshot (MIT licence). Source data: SuperCon database (NIMS) as distributed by Hamidieh (2018) via the UCI Machine Learning Repository, DOI 10.24432/C53P47 (CC BY 4.0) - not redistributed here.