EHA-DPRM proposes a fail-closed write-time governance architecture for retrieval-memory systems. Rather than treating external knowledge as a passive vector store, the framework specifies when information may enter memory, when atomic units may compose into higher-level entities, how neural retrieval may influence but not unilaterally authorize ranking, and when the system must HOLD or ESCALATE rather than generate. The architecture couples five mechanisms usually treated separately: write-time admission, compositional hierarchy formation, append-only provenance, bounded hybrid retrieval, and fail-closed answer authorization. Its central claim is not empirical superiority over existing RAG systems, but a formal reframing of retrieval memory as an admissibility-governed state-transition process. The proposal includes system-state definitions, admission gates, bounded scoring dynamics, EXACT1 decision semantics, invariants, MVP requirements, and an ablation plan for future empirical validation.
Abraham Rubinestock (Wed,) studied this question.