Governed Retrieval Augmented Generation with Guardrails Role Based Access Control and Audit Logging

Enterprise retrieval-augmented generation (RAG) must preserve source permissions while controlling unsafe content and retaining audit evidence. This paper integrates permission-tagged ingestion, authorization-aware retrieval, guardrails, and correlated logging, and evaluates the authorization boundary in a reproducible synthetic diagnostic. Five retrieval strategies were compared across eight corpus randomizations, each containing 750 documents, 12 principals, and 360 queries under five policy conditions. SQLite FTS5 supplied exact lexical ranking; no language model or Amazon Bedrock deployment was evaluated. Indexed filtering alone admitted unauthorized context in an average 99.5% of role-revocation queries and 57.7% of document-revocation queries. Adding a current-permission recheck produced zero observed context violations in the tested conditions. With 10% of private index metadata removed, authorized recall was 92.1%; policy-service outages forced all protected methods to abstain. An independent recomputation verified the saved retrieval and policy decisions. These results illustrate a known stale-permission failure mode and its mitigation under a trusted policy model. They do not establish production security, generated-answer safety, or regulatory compliance.