Governing Across Ontological Worlds

Abstract

Governing Across Ontological Worlds develops an architectural argument for enterprise AI systems that must operate across business meaning, governed analytical data, and material data systems. The paper begins from a distinction among three ontological worlds: the Business World of domain meaning and authority, the Data World of governed analytical identity and law, and the Material Data World of representation and computation. Its central claim is that integration becomes a governance error when these worlds are tightly coupled so that identity, meaning, predicates, measures, or authority established in one world automatically acquire force in another. The paper identifies three recurring forms of this failure: implementation capture, where a business concept becomes inseparable from its SQL or warehouse implementation; identity collapse, where one identifier is treated as one identity across business, analytical, and material worlds; and reference inflation, where shared reference is treated as permission to propagate predicates, measures, or claims. The governing deduction is that a crossing can be governed as a crossing only while the worlds on its two sides remain distinct. Once the distinction is erased, the transfer of authority still occurs but is no longer represented as an independently governable act. The alternative is explicit governed crossings: versioned, purpose-bounded relations among distinct endpoints with stated evidence, authority, permitted consequences, non-transferred authority, and revision rules. Such crossings can also create new shared meaning without requiring either endpoint to absorb the other: the relation may be new while the endpoints remain themselves. A worked Finance/Support example shows how an application can compose Billing Customer and Service Customer through an approved relation without creating one universal Customer identity, including a case where a plausible Revenue attribution must be refused because correspondence does not establish analytical permission. AI agents make the issue urgent because one probabilistic component can traverse interpretation, semantic resolution, analytical construction, material execution, interpretation, and action in one continuous process. The paper therefore proposes a compact design rule: let intelligence traverse broadly; let authority traverse only through governed passages. The argument is implementation-independent and does not require any particular ontology language, semantic layer, query language, or execution system.