From what I'm seeing, the ECMWF guys are feeding these models their initialization data to run them on a testing basis.. with good results.I mean modeling can only be as good as the data/obs put in to it at the start of the run. I'd assume that's the hang up for making any big jumps in med-long range verification.
A "perfect" model is still going get skewed off course by the observation data put in to it.
All I know is euro is getting rid of the op runs soon, in favor of the eps and they are working on their own machine learning model based off tech of some of these good ones out there.