Building an AI That Refuses to Say "Safe": Reliability Engineering for an Offline Disaster App