Can't do this YET one method to reduce this could be to: create a response to query, then before responding to the human, check if answer is insane by querying a separate instance trained slightly differently...

Give it time. We will get past this.