Another possible suggestion. Maybe choose random points that are within a set radius of points chosen along the borders? So perhaps choose first a random selection of points on the border, then choose random points within a circle (or perhaps just a square with a set delta in the lat/long) that are "nearby to the border" - then measure your error rates for those points at various boundary simplification tolerances? That'd remove the "middle of the state" random points where the border tolerance inevitable makes no difference.