I was trying to use transformpieces or a copy SOP to transform pieces to their corresponding transform matrix after simulating them with bullet.
When I apply some scaling to the points before simulating, they end up with slight offsets to what the dopnet result is, but only if the simulated pieces are fractured geometry, if it's spheres or boxes it seems to work fine.

Am I missing something here on how to do this properly?

I found the problem. I have to either apply the scaling before the rotation, which the copySOP or transformpieces does not do by default, or multiply both 3x3 matricies before setting the transform intrinsic.

I don't know if this is the most elegant solution, but if anyone is interested, here's the fixed file: