What we're learning building auditable compliance AI.
Language models are trained to complete, not to push back. Why that makes unchecked completion — not hallucination — the real compliance risk, and the planner–generator–evaluator loop that makes AI output gradable.
