Four questions nobody can answer, one architectural decision that answers all of them, and how a governed AI stack turns every model upgrade into evidence instead of risk.
