I think the problem with the SAS  shaking the plane apart is based in the dual use of large surfaces as both roll and pitch control. As soon as there is a little roll induced, those make overcorrections when having input in both axis (and there is nearly always some Pitch Input) that destabilise pitch and roll simultaneously, leading to the runaway of control.  I’ve reduced this problem by a large margin in just splitting those in two distinct surfaces, the big elevators solely for the