Complete procedures with a way back.

Build a production RAG evaluation program that separates retrieval recall and precision from answer faithfulness, citation validity and support, correctness, usefulness, and abstention; calibrates automated graders with humans; blocks security regressions; and carries evidence through regression, canary, continuous evaluation, and rollback.

Build an evidence-governed answer pipeline with authoritative sources, ACL-safe retrieval, claim-level citations, independent verification, empirical abstention thresholds, representative evals, and coupled rollback.