I am not sure I got you completely understood.
We actually do a full debayer using RED SDK, then we have an RGB image with several color pixels being "invented" by the RED SDK debayer. We throw those away and remain with those that have been seen by the sensor, so a G1 R B G2 representation for example.
Clearly this is not exactly the same as REDs internal RAW data, but as they closed that approach in the past, its the only way remaining. The good point is, that this virtual RAW image contains already final color rendition from the RED SDK, so its quite ok.
Still I recommend using RGB444 instead, except you really need smaller files or different deBayer filters (such as sharp for contours / keying, but soft for inner image data like skin).
Do you say you can bypass RED's deBayer to access their RAW data?
Thats not possible anymore, as to my knowledge with the SDK at all.
If you really can access the original RAW data prior to deBayer, then it should be a fairly fast process.
And what do you mean the image looks like a 0.5 gaussian blur being applied to the ROCKET output?
I'd say you might just witness the different quality between software deBayer and the ROCKET deBayer...