Tracing Cross-Field Responses to Input Perturbations in Transformer-Based Weather Forecasting Models

Aggregate forecast scores do not show how an error in one atmospheric field changes the rest of a multivariate forecast. We audit this behavior in two transformer-based weather models, Aurora and Pangu-Weather, by replacing each of 19 inputs with season- and UTC-matched climatology while holding the other fields fixed. Paired rollouts over 48 dates produce directed $19\times19$ response matrices at +6 and +24 h, supplemented by six-dose sweeps and grid-cell error maps. In both models, replacing $Z850$ changes $V850$ about nine times more than the reverse intervention. The diagonal $Z500$ displacement falls by roughly half from +6 to +24 h, while responses in other fields persist. Lower-level geopotential perturbations affect temperature, humidity, and wind, although terrain and below-ground extrapolation may explain part of this pattern. The audit can guide input-quality checks for forecasts used by early-warning services. It measures model responses rather than atmospheric causality; event-level benefits remain untested.