I've looked at 1&3-ply evals for money play/matches cubeless and cubeful and noticed the following.
If you start with either 0 ply or 2 ply as a basis (they're usually pretty close) and then ask for a 1 or 3 ply eval the equity is ALWAYS lower. (I couldn't find one higher)
It doesn't matter if it's positive or negative to start with.
Is that simply the way GNU's net behaves from training or is there more to it?