What Random Splits Measure in Polymer Glass-Transition Prediction: Chemistry Holdouts, Applicability Domains, and...

Predicting the glass-transition temperature (Tg) helps select polymers for a service-temperature window, but accuracy on familiar structures may not transfer to a new family. We compare validation protocols on 7367 curated homopolymers. Histogram gradient boosting gives R² = 0.899 and mean absolute error (MAE) = 23.9 K under random cross-validation, 0.857 and 28.5 K under scaffold holdout, a class-macro mean of 0.355 and 43.3 K under curator-class holdout, and -3.79 and 79.5 K under high-Tg extrapolation. Similarity explains only 6-11% of within-protocol absolute-error variance in the first three settings. A separate density functional theory (DFT) band-gap dataset shows a moderate fingerprint-cluster holdout penalty and poor property-range extrapolation for the tested baselines. Random-forest split-conformal empirical coverage falls from 0.899 under random splitting to 0.683 and 0.214 under class and property holdout; the tested prospective similarity gate provides little protection. A fold-fitted group-contribution prior is more sensitive to uniform additive bias in a squared-loss penalty than as a neural-network input, but this comparison is architecture-dependent. Class-holdout intervals and tests are descriptive or nominal because training sets overlap. The 279 multiply reported polymers support only the between-source variability analysis. Validation should match the intended polymer family and property range.