Isn't G a learned probability?
If you put something through a softmax the output is (trivially) a valid PMF. Does that matter? You're not sampling from it.
If you put something through a softmax the output is (trivially) a valid PMF. Does that matter? You're not sampling from it.