If there are path dependencies developing, like nonlinear metal plasticity or frictional sliding, one might get different end results (stepped vs ramped) with a large enough initial interference. Also, the contact stiffness updating algorithm might be producing different end values. Are the final contact penetration values different? One can track these values via the results tracker under Solution Information branch. If they are, one could try turning the stiffness updating off or switching to Normal Lagrange algorithm, but that might create convergence challenges. For the most accurate result, consider how the parts are physically put together. With large enough initial interference and path dependencies,try to model the actual assembly process instead of trying to reconcile the interference with the parts in their final position.