The National Centers for Environmental Prediction has been running in a delayed mode a multi-model Short Range Ensemble Forecasting (SREF) system since April, 2000. A version of this system is expected to begin running in a real-time mode in May, 2001 (Du and Tracton, this conference). This paper focuses evaluation of the ensemble for selected case studies (e.g., the January 25-26, 2000 east coast snowstorm) of significant weather events. Presented will be a discussion of the value of the ensemble output in the context of a "virtual" operational environment. Emphasis, however, is upon on diagnosis of the reasons for the differences between the relatively good versus bad performing ensemble members as a function of differences in initial conditions (analysis uncertainties and data problems) and the dynamic/thermodynamic mechanisms that lead to the growth of differences between the respective ensemble members. It will be shown that the "regional breeding" used to generate initial state perturbations produces divergence between forecasts (error growth) consistent with thermodynamic/dynamic expectations.