The tools are now available to make a clean comparison of the skill of ensemble forecasts using properly constructed Monte-Carlo perturbations (from an coupled ensemble/data assimilation filter) vs. singular vectors using an analysis-error covariance initial norm (using time-varying analysis and forecast-error covariances developed from the same large ensemble filter). Each is in their own right about the best random and the best singular-vector method that can be constructed. We will conduct this experiment using a large ensemble in a high-resolution Held-Suarez primitive-equation GCM under perfect-model assumptions. We will use standard ensemble forecast verification methods to compare the skill of probabilistic forecasts generated from each ensemble. This sort of clean comparison will hopefully shed some bright light on the question of whether dynamically constructed or random perturbation methods are preferable for ensemble forecasting.