Group-aware and external validation distinguish interpolation from transfer in machine learning for aqueous pollutant removal.

Machine-learning models are increasingly used to rank adsorbents and treatment conditions for removing hazardous contaminants from water. A random row split can, however, place measurements from the same material, pollutant, or source study in both training and testing, so a high score may describe interpolation rather than transfer. We surveyed 75 aqueous-removal studies and re-tested five reconstructable adsorption datasets while holding the rows, physically available features, learner, hyperparameters, and seed fixed. Sixty-four studies (85.3%) used validation that did not preserve a reported entity boundary. Replacing shuffled five-fold validation with leave-one-specified-group-out testing reduced pooled R² by 0.108-0.983. One benchmark retained R² = 0.878 under chemical-signature holdout, whereas a uranium-biochar compilation fell from 0.981 to approximately zero under source-study holdout. An independent model-data transfer test on 60 condition means from four adsorbents gave R² = -7.187 after source-study holdout had given 0.510. Because dose was absent from the published feature set and external descriptors and laboratory procedures differed, this external result does not isolate splitting as the cause. A validation ladder that reports row-wise interpolation, specified-group transfer, and independent external performance makes the decision boundary of an aqueous-removal model explicit.