What happens when an AI can identify a commitment—but shouldn't be allowed to make it authoritative? I built a meeting intelligence workflow where memory, inference, evidence, and human approval stay separate. Here's how I designed that boundary.