maybe it's trained on non-published documentation pages?