Reproducibility
Code, data, and environment pinned well enough that the exact result can be regenerated — not approximated, regenerated. A model that only ever ran once, on its author's laptop, with whatever package versions happened to be installed that week, isn't a production asset yet. It's a demo that hasn't failed in front of anyone else.
This is what makes debugging a regression, auditing a decision, or safely retraining possible at all. Skip it and every later stage inherits an artifact nobody can fully explain.