Ahh, I tried so hard to get people to be precise on that thread about what they wanted. @andrewgelman asked for that here: here and here and I didn’t see any objections or clarifications. It also seemed to match up with the storage table, where iterations are changing most slowly. Does that mean the table representing the internal storage is also not what you wanted?
How do you propose having fit$theta return 3x2x4x1000 or 3x2x4000 depending on whether someone intends to use it for convergence or inference?
@andrewgelman, would you be happy with the ordering Ben is proposing?