A research team has built a compliance review system that stops hiding its reasoning inside a black-box prompt.
PolicyGuard is a neuro-symbolic framework that converts organizational policies — NDAs, negotiation playbooks, internal guidelines — into typed relational logic rules. During a review, a large language model answers narrow, targeted questions about specific document passages, and a separate symbolic evaluator applies the formal rules to flag violations. The system was tested on NDA compliance, checking contract clauses against company-specific negotiation policies. Crucially, the policy logic is explicit code, not an implicit blob embedded in a prompt.
This matters because the dominant approach to AI-assisted compliance is to stuff the policy and the document into a prompt and hope the model applies the rules consistently. That works well enough in demos, but it is nearly impossible to audit, update, or test systematically — every policy change requires a new round of prompt engineering and eyeballing. PolicyGuard's separation of formalization, interpretation, and evaluation gives legal and compliance teams something closer to a test suite than a magic eight ball.
Neuro-symbolic AI has been a recurring academic promise for years, often stalling because the symbolic layer is brittle or the integration overhead is too high; whether PolicyGuard's approach holds up outside controlled NDA experiments is the open question.