Cross-study shift defines the reliability limits of FTIR polymer classification

Automated Fourier-transform infrared (FTIR) classifiers for microplastics are often evaluated by randomly splitting spectra pooled from several studies, although deployment requires transfer to new sources. We harmonized 4,783 labelled spectra from four independent studies and compared a study-mixed random split with leave-one-study-out (LOSO) evaluation. We then examined task composition, source-associated structure, nested deep and domain-oriented models, external reference-library pooling, and selective prediction on locked domains. XGBoost achieved random-split macro-F1 of 0.972 but equal-study present-class LOSO macro-F1 of 0.672. The gap was class- and source-dependent: PE/PP transfer remained strong, whereas adding PET and PS reopened most of the shared-fold deficit despite adequate class support. Polymer class dominated the global geometry of the harmonized data, yet source study remained highly predictable, including within PE and PP. Position-aware convolutional, self-supervised and domain-adversarial configurations did not consistently outperform XGBoost, and external libraries produced non-monotonic effects, indicating that directional compatibility mattered more than library size. A maximum-probability threshold selected only from inner study holdouts identified lower-risk subsets on locked De Frond and Open Specy domains, although coverage was domain-dependent. These results support a practical trust framework for multi-source FTIR classification: validate by study, report fold- and support-aware metrics, screen external libraries for compatibility, and defer spectra outside the model's reliable operating region.