Build a DeepSeek prompt regression suite with paired cases, strict output checks, critical failure gates, and a reproducible offline Python checker.
Build a DeepSeek routing policy around capability gates, task-specific evaluation, accepted-response cost, bounded alternatives, and safe release controls.
Choose DeepSeek thinking mode and reasoning effort with matched tasks, quality gates, honest latency metrics, and a tested offline evaluation helper.
Evaluate DeepSeek RAG with an offline Python fixture, evidence coverage, citation checks, calibrated judges, safety tests, and reproducible release gates.