Models hallucinate more than you think

What are your favorite observability tools and techniques for verifying output?