Ensemble Kalman filters are increasingly used to estimate and predict the state of the atmosphere and oceans. The models used are under-resolved, which leads to three types of errors: numerical errors, subgridscale errors, and representation errors. The state of the system can be separated into a large-scale/resolvable part and a subgrid-scale part. Numerical errors are associated with errors in the discrete representation of the large-scale part; subgrid-scale errors are associated with errors due to the interaction of resolved and unresolved scales; representation errors are associated with the mismatch between what is being observed (the full system state) and what is being modeled (the large-scale part of the system state). There are many approaches to reducing, and accounting for these errors in a filtering scheme; this talk explores three of them – covariance inflation, stochastic subgridscale parameterization, and model numerics – in the context of an idealized fluid system: two-layer quasigeostrophic dynamics.