T-RKG: A Knowledge-Based System for Cross-System Regulatory Conflict Detection in Enterprise Records Governance

Abstract

Enterprise records governance is a coordination problem across systems that share no unified regulatory knowledge. Email platforms enforce communication-retention, ERPs enforce financial regulation, document repositories enforce their own schedules, and no system sees the others’ applicability logic. Conflicts arise wherever two applicable regulations impose contradictory requirements on the same record. The framing problem is sharper than fragmentation alone: a single transaction can carry tax, privacy, and audit obligations from three jurisdictions, each conditioned on a different attribute that lives in a different source system, so any architecture that governs only content can miss the conjunction. This paper presents T-RKG (Temporal Records Knowledge Graph), a knowledge-based system with three contributions: a formal governance ontology of 47 domain classes and 23 object properties, plus a six-class temporal-relationship submodule (53 OWL classes in total), delivered in OWL/Turtle as supplementary material, that encodes records as composite governance entities; a typed temporal relationship model in which hold propagation is governed by the declared semantic properties of inter-record connections; and a conflict-detection mechanism that applies ontological applicability reasoning, including a defeasibility axiom for GDPR Article 17(3) exemptions. The substantive empirical content is two-fold. First, on per-record applicability measured against a held-out clean-label snapshot across ten seeds and six scales (1K to 100K records), T-RKG attains F1 of <inline-formula> <tex-math notation="LaTeX">$0.826 \pm 0.011$ </tex-math></inline-formula> under 10% per-attribute label noise, against <inline-formula> <tex-math notation="LaTeX">$0.604 \pm 0.035$ </tex-math></inline-formula> for a Siloed baseline and <inline-formula> <tex-math notation="LaTeX">$0.141 \pm 0.011$ </tex-math></inline-formula> for a No-Ontology baseline; perturbing the labels moves the score, which is what distinguishes a genuine result from a tautology produced by the clean-regime F1 of 1.000. Second, a SHACL Core baseline — the strongest expressive baseline within standard RDF validation — detects roughly 4% of T-RKG’s conflicts at 100K records and takes about <inline-formula> <tex-math notation="LaTeX">$62\times $ </tex-math></inline-formula> longer to do so; the rule families it cannot express (temporal-interval intersection, defeasibility, cross-record propagation, Art. 17(3) exemption) are catalogued with their W3C-specification gaps in the supplementary material. The total-conflict comparison against attribute-blind baselines (T-RKG 84 to 7,610 conflicts per dataset; both baselines zero cross-domain) is reported in Table 6 as a baseline check whose outcome Corollary 1 forces by proof, so it is not the central evaluation result. Hold propagation under a Conservative policy expands the hold set <inline-formula> <tex-math notation="LaTeX">$3.07\times $ </tex-math></inline-formula> over direct seed identification and runs roughly <inline-formula> <tex-math notation="LaTeX">$39\times $ </tex-math></inline-formula> faster than a flat-list and <inline-formula> <tex-math notation="LaTeX">$175\times $ </tex-math></inline-formula> faster than a SQLite recursive-CTE baseline on identical workloads on this hardware; the absolute multipliers will reshape on different hardware, though the qualitative ranking holds.