Funnily enough, with the most recent models (having reduced sycophancy), putting in the wrong assumptions often still leads to the right output.