Abstract The stability of machine‐learning models is generally evaluated by their ability to tolerate input noise. This study suggests that where the noise enters the physical trajectory is equally important. We predict subsurface soil moisture across depths of 10–40 cm from 5 cm observations at 21 globally distributed international soil moisture network stations by using lag‐aware Fourier Neural Operator. We apply perturbations separately to critical and non‐critical hydrological states to isolate state‐dependent model fragility. Critical states maintain the data‐efficient role, allowing reduced training subsets to recover the 70% baseline skill. However, the same states are also fragility points: perturbations applied to critical states produce stronger degradation, lower transition thresholds, and higher collapse probabilities than equal‐amplitude perturbations applied to non‐critical states. We find that the model breakdown is not governed by noise magnitude alone, but by the interaction between input uncertainty and physically informative states in the learned hydrological mapping.