handling. What success looks like AI outputs are grounded in data, and their supporting sources can be inspected wherever the risk tier requires it. Task, retrieval, citation and workflow quality are measured through repeatable evaluation and regression suites, with hooks exposed for independent evaluation. Prompts, models, orchestration logic, datasets … step workflows with explicit human review, approval, escalation and safe-failure (abstention) paths. Work with stakeholders to create evaluation datasets and automated tests for task success, retrieval quality, source support, robustness, latency, cost and regression. Uphold security-conscious coding standards, building robustness against prompt injection, unsafe tool use, access ...