Thanks for the link a1ex, that's the most helpful thing I've read. I think I've started to make some progress in my understanding. Please correct me where I'm wrong:

Compute XYZKelvinWB matrix using the "chromatic adaptation" techniques in the Bradford cone domain (I'm pretty sure now, that I've got this part right). Then use the XYZKelvinWB matrix to transform "white" (1,1,1) to some different color temperature (I assume this is what the XYZKelvinWB matrix does, but I'm not totally sure). Then convert this new "white" back to the camera raw color space using the inverse of the "ColorMatrix1" (which if I understand correctly this matrix is a Cam->XYZ). Now we have a neutral color in camera raw coordinates right? So then the multipliers (aka AsShotNeutral) are simply {R/G, 1, B/G}. Mathematically this would be:

CamNeutral = [ColorMatrix1]^{-1} * ([XYZKelvinWB] * [1,1,1])

AsShotNeutral = [ CamNeutral _{R} / CamNeutral _{G}, 1, CamNeutral _{B} / CamNeutral _{G} ]

Is this correct, or am I totally off?