Centaurian Architecture: A Quantum-Cognitive Architecture for Interpretable, Embodied AI with Neural Periphery

Modern large language models achieve remarkable linguistic fluency but remain opaque, resource-intensive, and fundamentally uninterpretable — every output is a stochastic sample from a trillion-parameter black box. This paper presents the **Centaurian Architecture (CA)**, a multi-layered system for human-like AI that preserves the interpretability and traceability of symbolic cognitive modeling while selectively incorporating lightweight neural components exactly where they demonstrably outperform rule-based alternatives. A critical preliminary clarification: the architecture's Quantum Personality Model (QPM) is an instance of **Quantum-Like AI (QLAI)** — it employs the mathematical formalism of quantum mechanics (Hilbert spaces, density matrices, unitary evolution) as a modeling language for cognition, running entirely on classical hardware. No quantum computer is required. The architecture integrates four core subsystems: (1) a **QPM** encoding the Five-Factor Model of personality into a 12-qubit Hilbert space formalism with empirically calibrated entanglement parameters and operationalized Lindblad decoherence dynamics; (2) **small language models** serving as linguistic transducers (Qwen2.5-7B-Instruct standardized across both deployment tiers; 7B is the empirically validated minimum, established by Experiment 1, with 3.8B incoherent on JSON-based SCI prompts) that convert structured cognitive outputs into natural language without performing reasoning; (3) **custom JA/LI procedural facial animation** synchronized with lightweight neural text-to-speech; and (4) a **domain-specific knowledge architecture** combining RDF/OWL ontologies with embedded vector retrieval for grounded, hallucination-resistant content generation. We provide complete specifications for the quantum circuit design, the formal QPM-measurement-to-SLM translation protocol, the ontology schema and vector retrieval pipeline, LoRA fine-tuning methodology, edge deployment resource estimates, empirical validation framework (Experiments 1 and 2), the Self-Model Component architecture (Section 17), and an extension path from software-only virtual agents to physically embodied humanoid robots. The system's total memory footprint is ~5 GB (base CA, Tier 1) to ~6 GB (SMC-enabled, Tier 2 with the LoRA-10K SCI grounding adapter), enabling deployment on current edge hardware (NVIDIA Jetson Orin, Qualcomm Snapdragon 8 Elite, Apple Silicon) while maintaining end-to-end traceability — every behavioral decision is auditable from situative input through quantum state evolution to observable output, with neural components confined to bounded I/O transduction roles.