What are the primary failure modes you anticipate when deploying generative models in a laboratory or government setting?
