Neural Orchestry Modular Unite (NOMU): A Deterministic Cognitive Architecture with LLM as Periphery for Personal Devices

Personal cognitive assistants and mobile autonomous agents frequently rely on large language models (LLMs) as central reasoning engines. However, this architectural paradigm introduces non-determinism, unpredictable execution latencies, opacity in decision provenance, and severe security risks regarding unconstrained system actuation. In this paper, we introduce NOMU, a deterministic, platform-agnostic cognitive architecture operating natively on personal edge devices. NOMU decouples high-level cognitive evaluation from non-deterministic external interfaces by enforcing the core foundational principle: “The core thinks; peripherals translate; the LLM never decides.” All incoming signals undergo a rigorous four-stage state pipeline (propose → validate → authorize → execute), governed by 24 machine-verifiable architectural invariants and a three-axis epistemic trust model. We present a full empirical evaluation on an edge deployment (Android/Termux, Python 3.13.13), demonstrating a median pure cognitive latency of 4.53 ms, O(1) knowledge base search time up to 10,000 concepts (13.87 µs), a 47.4% comprehension rate and zero cognitive failures on a calibrated 95-phrase corpus (the corpus was constructed from the vocabulary declared by the system; the primary metric is the comprehension rate, not the 100% match with the expected label).Adversarial string inputs up to 10,000 characters produce zero unhandled errors. Our latency breakdown reveals a think-act ratio of ∼267× between pure core cognition (4.53 ms p50) and the latency of turns that actually execute a real action (1238.93 ms p50, measured exclusively over turns classified as RESULT). We analyze system sustainability over 10,000 continuous execution turns and document structural memory dynamics.