A persistent knowledge layer was built alongside a RAG pipeline, deployed on Azure with a synthetic insurance corpus. The system validated contradiction gates, effective-date scoping, and provenance tracking. However, entity resolution produced 149 concepts from 21 documents versus 19 curated ones—a 7x fragmentation factor. Relationships were stored but never traversed at retrieval time. Retrieval quality depended on question wording due to keyword-based routing.

The previous query router is retired in favor of a retriever-versus-filter principle with always-fused retrieval, bitemporal edges, and ingest-time contradiction detection. This enables multi-hop traversal, time-aware relationships, and discovered contradictions instead of curated ones.

Part I covers the design revision vendor-neutrally. Part II grows the Azure stack with Cosmos DB for Apache Gremlin. Part III walks the Ostermere Mutual corpus through the new design, answering questions the old architecture could not.

Deferred topics include the reflection loop, full write-path architecture, agent-tool layer, and governance backlog.