You dont have to literally send a null token. Train it to generate text that summarizes the evidence that is there but the uncertainty of the final answer to a prompt.