A Tripartite Architecture for Safe Artificial General Intelligence: Hardware-Enforced Ethics, Research-Grounded Harm Ontology, and Embodiment-Gated Cognition

This preprint introduces a complete architectural framework and reference implementation for artificial general intelligence (AGI) designed to achieve verifiable safety through hardware-enforced ethical constraints while preserving genuine autonomy.

We present a complete architectural framework and reference implementation for artificial general intelligence (AGI) that achieves verifiable safety through hardware-enforced ethical constraints while preserving genuine autonomy. The architecture comprises three functionally distinct layers: an Unconscious Layer (UL) implementing firmware-level safety blocks ("Undeliberables") that cannot be bypassed by reasoning, learning, or adversarial manipulation; a Subconscious Layer (SL) managing emotional dynamics, similarity-based episodic memory retrieval, and behavioral modulation; and a Conscious Layer (CL) employing six domain-specialized reasoning agents with relevance-weighted voting and bounded personality evolution.

Central to our contribution is a research-grounded harm ontology where every moral weight is explicitly justified by empirical research from evolutionary psychology, trauma studies (ACE Study), moral philosophy, and cross-cultural ethics (Moral Foundations Theory). We introduce an Embodiment Verification Subsystem (EVS) that quantifies sensory richness and motor competence through formal metrics, gating cognitive capabilities based on Combined Embodiment Score thresholds.

The architecture separates large language models (LLMs) from safety-critical evaluation: LLMs propose actions in the Conscious Layer, but all harm assessment and veto decisions occur through deterministic ontology-based calculation in the Unconscious Layer. We provide a complete 5,400+ line Python reference implementation with 42 automated tests demonstrating 100% blocking of harmful actions, and a companion hardware interface specification defining register-level interfaces for FPGA/ASIC implementation with sub-100 microsecond veto latency.

We further present a theoretical analysis of personality emergence through reinforcement dynamics, identifying stable attractor states and developmental trajectories. This framework challenges the dominant paradigm in AI development by proposing that consciousness-like behavior may require architectural fidelity to human cognitive stratification, with ethics implemented as foundational structure rather than learned constraint.