AquaContam: machine-learning models of drinking-water contamination learn who is monitored as much as where contamination occurs

Tyler James Newton 

Machine learning informs drinking-water monitoring, yet administrative data record the
monitoring process as much as the contamination. Using AquaContam (over 4 million samples
from 95,223 U.S. water systems and ambient monitoring locations), we find provenance proxies
dominate feature attributions, and an apparent lead-model AUROC of 0.962 is largely a target-
leakage artifact (corrected skill 0.699, 0.550 without provenance). A provenance-reduced PFAS
signal survives (in-region AUROC 0.785) but transports modestly (AUROC 0.691, 95% CI
0.615–0.769), and spatial leakage separately inflates baseline-feature AUROC 0.699 → 0.869.
Monitoring is itself inequitable: communities of color are sampled 1.85× more intensively
(attenuated but significant under source adjustment), with heterogeneous detection burden
(1.48, 95% CI 1.24–1.88; significant in 3 of 10 regions under a spatially-aware null), whereas
low-income communities are under-sampled (0.78×). Although the group false-negative gap
(0.50 versus 0.44) is not significant, the model is worse-calibrated for high-people-of-color
systems (calibration error 0.18 versus 0.06). AquaContam provides a reproducible two-
confound protocol with one converged detection task and documented boundary tasks.