Definition
Graph-native data governance attaches permission, policy and purpose to nodes and edges, so an agent can only traverse what it is allowed to see. The path is the control, not a later filter.
The problem
Most enterprise controls sit around systems: a warehouse policy, a document ACL, a prompt instruction. Agents assemble answers across all three. If governance is applied after retrieval, the agent has already seen what it should not. If it is applied only in the prompt, it is a suggestion.
What exists today
- Role-based access on source systems, which does not survive a join across systems.
- Retrieval filters on document chunks, which cannot express relationship-level permission.
- Prompt-level policy, which is not enforceable.
Deca10 research question
Can permission and policy be evaluated as graph traversal constraints, such that two agents asking adjacent questions receive different paths — and that difference is inspectable?
Evidence
The evaluation bar is path-level deny/allow correctness, not document recall. Work in progress: agent path traces against a governed organizational graph.