Systemic Knowledge Compilation: Operationalizing n-ary Hypergraphs for Local LLM Reasoning

Abstract

As Large Language Models (LLMs) continue to scale, the transition from monolithic, opaque information storage to structured, queryable cognitive architectures becomes imperative for the development of Proto-AGI systems. Current Knowledge Graphs, restricted by binary edge constraints, fail to capture the multidimensional semantic interactions inherent in complex domains. This paper introduces a computationally sustainable architecture for the construction of n-ary mathematical hyperontologies, operationalized through a modular, local extraction pipeline. We present intuyt: macro, a custom fine-tuned Mixture of Experts (MoE) model with 3 billion active parameters, designed specifically for domain-specialized topological extraction in Mexican Spanish and indigenous language corpora. By implementing a greedy, token-bounded partitioning algorithm and deterministic JSON sanitization, we enable the compilation of hyperontologies exceeding 20,000 nodes on edge hardware, while bypassing cloud-based latency and privacy risks. Finally, we propose a WebGL-based visualization framework employing selective rendering and topological bounding $(N_{limit}=300)$ to enable fluid, real-time exploration of hyper-dimensional memory structures. Our results demonstrate that local, iterative cognitive loops achieve superior structural adherence and cultural fidelity compared to general-purpose cloud-based APIs, establishing a viable pathway toward efficient, persistent, and self-governing Proto-AGI architectures.