OK, this is kind of terrifying, but I believe that this is how I do it:
p model['AND Output'].transformation.backward(np.array([ 0.2386266 , 0.13873085, -0.53425323, -0.16257509])).eval()
Where the argument to backward comes out of the debugger output above.
That does for transformed variables, but doesn’t get me to the deterministic ones, which would really make the debugging easier.