Abstract—Knowledge graphs play a vital role in grounding
Large Language Models (LLMs) in verifiable facts, which matters
wherever facts evolve continuously and stale information is costly,
as in clinical informatics. However, existing Retrieval-Augmented
Generation (RAG) systems suffer from “structural hallucination,”
where contradictory “ghost nodes” corrupt retrieval context.
In order to resolve conflicts at ingestion and eliminate stale
graph topologies, it is crucial to enforce deterministic symbolic
governance over probabilistic LLM outputs. For this purpose,
a neuro-symbolic framework, Arbiter-RAG, is proposed which
integrates a deterministic arbitration layer with a temporal
knowledge graph. Incoming predicates are classified into four
types i.e. Immutable, Temporal, List, and Functional, each with
a fixed resolution strategy. On the controlled generator, the full
system yields F1=1.000 against 0.802 for the baseline; these values
are logic-validation checks rather than empirical clinical accuracy.
On the MIMIC-IV-on-FHIR Demo, 875,609 facts from 100 patients
are materialised with zero duplicate active keys and zero invalid
valid-time intervals. At 20% valid-timestamp late arrivals, event-
time arbitration has no F1 loss while arrival-order arbitration
falls by 0.067. The event-time and arrival-order systems lose the
same F1 under incorrect values, units, patient references, and
predicate mappings, so no content-integrity advantage is claimed.