A Right Response Is Not a Learned Law: What Perturbation, Constraint and Interpretability Tests Show About Whether Data-Driven Atmospheric Models Learn Chemistry and Dynamics, Why a Constraint Must Be Expressible Before It Can Be Missed, and the Evid

(c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. Life of Research Weather foundation models fine-tuned on reanalysis now issue global air-quality forecasts that match or beat a chemical transport model on most verification targets. A July 2026 study of one of them, Aurora, reports that it reproduces a first-order ozone response to added nitrogen oxides, yet produces negative concentrations and chemically inconsistent species combinations, relaxes wildfire plumes toward background, and carries internal features that steer the forecast without mapping onto atmospheric processes. In this analysis the author reviews and synthesises published results to show that the question such studies ask - has the model learned the process, or exploited regularities in its training data - bundles four findings that published evidence separates: skill against a reference, the response to an intervention, the satisfaction of a constraint, and an internal variable that mediates a process. Skill coexists with distorted emission responses and with budget drift. A correct-looking response can arise from numerical noise or from regional errors that cancel, and a missing one can be the price of training on coarse-grained data. Internal features can be causal yet entangled across species, or can detect a phenomenon without driving it. The paper argues that a constraint must be expressible in a model's state before its violation says anything about learning: positivity is expressible in the anchor model's five trace gases, while a nitrogen atom balance is not. It consolidates 30 published values in a table and a figure, states a procedure for reading a claim that a model has learned a process, and names six measurements that would settle what the anchor study leaves open.