arXiv:2611.25818v1 Announce Type: new Abstract: Pipelines that let a large language model (LLM) make discrete decisions need to know which decisions they may act on unattended, and a record that explains each one.