NV Summarize the deployment risk for the new inference endpoint. AI The highest risk is the cold-start latency on the first request after scale-out. Keep the rollout limited to the staging pool until p95 startup time stays below the service target for three consecutive runs. NV What should the operator check first?