Oh I see. And nice approach to getting the signed difference. That looks to me like the right thing to be doing then.
Good point about arccos only giving half the range. There might be a way to extend the approach from my last post to get the sign right as well, but since all that matters is the input-output behaviour of this function (i.e., target points for the decoders to fit, with respect to the sampled eval_points) what you have here should be good!
Let us know if the model is still struggling to perform and we can look into some other possibilities for improvement.