AN1: A Dual-Substrate, Non-Statistical Architecture for Deterministic and Adaptive Intelligence

This technical note presents AN1, a dual-substrate, non-statistical cognitive architecture designed to separate ambiguity handling from executable reasoning. AN1 decomposes cognition into two distinct computational substrates: a Hypothesis Formation Engine (HFE), which generates candidate interpretations for noisy, ambiguous, or novel inputs by operating over structured representational spaces (e.g. geometric, symbolic, or relational), and a Causal Verification Engine (CVE), which operates on an explicit causal graph to perform logic, planning, memory mutation, counterfactual simulation, and safety enforcement through exact symbolic execution. Unlike transformer-based or neuro-symbolic systems, AN1 does not implement reasoning, memory, or safety inside probabilistic models. Hypotheses are non-authoritative and are committed only after causal verification. Invalid or unsafe states are structurally unrepresentable at the execution layer, making hallucination impossible at the point of action. A deterministic arbitration boundary governs transitions between hypothesis formation and causal execution. This boundary does not learn, predict, or route between models. It enforces when ambiguity is permitted and when commitment is required, ensuring that intuitive or heuristic proposal mechanisms never directly produce irreversible action. The HFE is architecturally non-authoritative and may be instantiated by non-statistical abductive engines or, where practical, by statistical models used strictly for hypothesis generation. Transformer and other statistical models may therefore be integrated as peripheral tools for perception, language interpretation, or expression, but they do not participate in decision-making, memory persistence, or safety guarantees. The paper details the architectural design, execution model, and validation methodology, and reports results from a 33-protocol internal benchmark suite covering one-shot causal learning, identity persistence, counterfactual reasoning, infinite effective context, safety enforcement, architectural composability, and computational efficiency. Comparative analysis is provided against transformer-based agents, compressed-state architectures, and reactive retrieval systems. AN1 is presented as a replacement for statistical reasoning cores in agentic systems, while operating on a fundamentally different computational substrate. This document is intended as an architectural disclosure and technical reference for researchers, engineers, and system designers exploring non-statistical approaches to high-assurance, adaptive artificial intelligence.