Spatial validation reveals transferability and allocation tradeoffs in national PFAS screening using UCMR5 data.

The 2024 U.S. PFAS National Primary Drinking Water Regulation has heightened the need to allocate limited monitoring effort under uneven data coverage and system capacity. Using UCMR5 records as of January 16, 2026, we evaluated screening models under spatial holdout and converted their outputs into descriptive prediction-set diagnostics and illustrative monitoring rankings. The analytical cohort comprised 23,331 sampling locations across 9436 public water systems (PWSs), with any-event PFAS detection as the primary outcome. We audited all 43 candidate predictors and excluded 25 current-snapshot constructions with possible post-sampling information; the primary reduced model used 13 predictors available before or at the first sampling event. Mean AUROC declined from 0.741 under grouped non-spatial cross-validation to 0.625 under 16-block spatial holdout. Across geographic specifications this gap decreased monotonically, from 0.161 with fine-grained administrative labels to 0.116 for the reduced model and 0.065 with no geographic labels. Under spatial shift, label-conditional conformal prediction reached positive-class coverage of 0.807 at α=0.10, below the nominal 0.90 level and without formal guarantee. At a 20% PWS budget, a hybrid probability-population ranking gave the highest observed PWS discovery (0.298), whereas population-first ranking reached the greatest population coverage (0.769) but selected no positive small systems. Size-stratified ranking increased small-system discovery to 0.303 but lowered population coverage to 0.271. Screening models should therefore be audited for temporal provenance and evaluated under spatial holdout, and allocation objectives specified, before model scores are used. These retrospective comparisons stress-test monitoring decision support within the observed UCMR5 cohort.