I review one AI agent or LLM workflow using public-safe information only, create 8-12 eval cases, map tool boundaries, identify top failure modes, and return an eval CSV, failure-mode register, release decision log, short reliability report, and practical next-step repairs.