My (poor) understanding was that the control run is based on the actual data collated at 00z, 06z, 12z, and 18z and pugged into the compuer forecasting model. These data are then pertubated a little and these also run (albeit at a lower resolution i.e. quicker). The forecasters then use their experience to judge which of the pertubations is most likely to be true and this becomes the operational run.
Please correct me if I'm really off the wall with my understanding.