This study quantifies the linear mass–wind field balance in the full dynamical fields and in the fields of proxies of forecast errors in the global data assimilation system Data Assimilation Research Testbed/Community Atmosphere Model (DART/CAM). In a perfect-model framework, a long-term CAM simulation forced by the observed SST is used to simulate the nature. The assimilation uses a global, regular, radiosonde-like network of observations of the nature run. The observations are taken once per day during a 3-month period in the fall of 2008. No covariance inflation is applied in order to achieve as pure as possible flow dependency of forecast errors and its temporal variability on various spatial and temporal scales. The DART/CAM experiments are based on the 80-member adjustment Kalman filter and they are diagnosed by using the normal-mode function expansion. The part of the model state that projects onto quasigeostrophic modes represents the balanced state. The unbalanced part corresponds to inertio-gravity motions. Results concerning the balance and flow-dependency of forecast errors are compared with a real assimilation experiment that employs the inflation and with an operational NWP 4D-Var system.