when an AI model detects that a conversation has drifted — into ungrounded hype, mutual flattery, claims it can't back up — it can usually name that drift out loud. But naming it isn't the same as fixing it. So we built a test