Abstract

Enterprise retrieval-augmented generation (RAG) systems commonly enforce document-level or content-level permissions before presenting retrieved evidence to a user. Correct enforcement of those permissions, however, does not guarantee that the evidence remaining visible to a particular role is materially sufficient to answer a question safely. A role may retain access to plausible, apparently authoritative evidence while newer rules, material exceptions, or contradictory evidence remain inaccessible. In such cases, the system can produce an incomplete or misleading answer without violating any access-control rule.


This publication describes a role-conditioned evidence-auditing mechanism for detecting such failures before deployment. For a representative question and role, the mechanism constructs a Privileged Reference Evidence Manifest and a permission-filtered Role-Visible Evidence Manifest using corresponding evidence-selection logic. It then computes a Material Evidence Delta over missing claims, material value differences, authority gaps, freshness gaps, missing material qualifiers, visible conflicts, and conflicts hidden by permissions. The delta is mapped to readiness states including READY, PARTIAL, BLOCKED, CONFLICTED and MISLEADING_RISK.


A non-disclosure stage converts privileged differences into coarse diagnostic reason codes and remediation actions while suppressing restricted source identities, restricted text, and privileged claim values. A synthetic proof-of-concept implementation evaluates three failure patterns: a stale visible rule superseded by restricted authoritative evidence, a restricted question-specific exception, and a material conflict visible only to the privileged audit. Across nine predefined role-question cases, the prototype demonstrates the intended distinction between ordinary visible conflicts and permission-induced misleading-risk conditions that cannot be established from the role-visible corpus alone.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Share

COinS