Am I corect in understanding that these 

Yes, see the link above.

It seems to be an ARW WRF, but has the advantage over the plethora of high resolution models initialised from the global models in that it includes an assimilation process for weather observations.

DuncanR wrote:

Yes, the assimilation piece intrigues me.  How do you integrate individual point observations in an x-terabyte modelling process?  By modifying the inputs in the following hour model run?  How could one do that on a systematic basis?

as Nick has mentioned.., the global models assimilate quite a lot of observational data - everything from satellite data to weather balloons, to ship reports. you can read about the GFS assimilation process here:
https://www.ncdc.noaa.gov/data-access/model-data/model-datasets/global-data-assimilation-system-gdas

 It is thought that one of the reasons the ECMWF outperforms the GFS is because of a superior data assimilation system.
Anyway, I have used the HRRR a fair bit over the years on LIS, and while it does sometimes accurately predict some small scale processes - like the little sea breeze cell that develops on the CT shore.., it seems  that it can be completely wrong sometimes too. 
In particular, with respect to the question of initialization.., it doesn't _only_ use observational data, and i frequently see that f0 (not exactly the initialization, but it should be very similar) is nothing like the actual conditions on the sound at the time- typically the wind direction is way off.