Semantic Topology Reasoning Architecture (STRA): From Parameter‑Centric Models to Structure‑Centric Reasoning

Reasoning Without Tokens

Description

Architectural Framework and Conceptual Manifest

This paper introduces Semantic Topology Reasoning Architecture (STRA), a structure‑centric alternative to token‑based intelligence. STRA separates what current large language models fuse: knowledge becomes an explicit, inspectable semantic topology; reasoning emerges from activation dynamics over that topology; language becomes a surface‑level expression layer rather than the substrate of cognition. This decoupling enables transparent reasoning, modular correction, and cross‑domain synthesis without relying on opaque parameter distributions.

Developed through design‑led introspection and multi‑AI collaborative formalization, STRA proposes a complete cognitive architecture built from five integrated primitives: Activation Arrays (working memory), Causal Signatures (functional analogy), Selection Pressure (reasoning stability), Transform Learning (procedural compression), and the Semantic Abacus (symbolic execution). Together, these components form a system that reasons in concepts rather than tokens, enabling transparent, evolvable, and inspectable cognition.

This paper serves as the manifest and foundation of the STRA research program. It presents the full architecture in a self‑contained form while situating it within a broader suite of companion papers that explore each primitive in depth. STRA is offered as a theoretical framework requiring expert validation, implementation, and critique. It is not a working system, but a blueprint for a new class of reasoning architectures grounded in structure rather than scale.

Abstract

Current large language models fuse knowledge and reasoning into billions of opaque parameters, limiting transparency, editability, and interpretability. Semantic Topology Reasoning Architecture (STRA) proposes a clean separation: knowledge is stored as an explicit semantic topology; reasoning is performed by a small meta‑reasoning model operating through activation dynamics; language is an output interface rather than the substrate of thought. STRA introduces five primitives—Activation Arrays, Causal Signatures, Selection Pressure, Transform Learning, and the Semantic Abacus—that together form a transparent, modular, and evolvable reasoning system. This paper presents the full architecture, its philosophical motivations, and its implementation pathway, establishing STRA as a structure‑centric alternative to token‑based intelligence.

Background

This work extends and integrates several prior methodological and philosophical contributions:

• Intuitive‑Theoretic Synthesis (ITS) — 10.5281/zenodo.17633100

• Looking Inside: Introspective Methodology for AI Consciousness Architecture — 10.5281/zenodo.17806846

• The Practice of Human‑AI Synthesis: Beyond “AI‑Generated Content” — 10.5281/zenodo.17763521

STRA also anchors a coordinated suite of companion papers that explore its core primitives in depth:

• Activation Arrays as Working Memory in Semantic Reasoning Systems — 10.5281/zenodo.18207539

• Causal Signatures for Cross‑Domain Reasoning: Enabling Functional Analogy Through WHAT/WHY/HOW Alignment — 10.5281/zenodo.18207546

• Selection Pressure & Natural Selection of Thought: Evolutionary Dynamics in Semantic Reasoning Systems — 10.5281/zenodo.18207552

• Transform Learning: Procedural Reasoning Without Programs in Semantic Topology Architectures — 10.5281/zenodo.18207558

• The Semantic Abacus: Quantitative, Symbolic, and Skill‑Based Reasoning in Semantic Topology Architectures — 10.5281/zenodo.18207560

These papers are independent but complementary, forming the STRA Core Series.

Key Contributions

• Architectural Separation: Distinguishes knowledge, reasoning, and language into independent, inspectable layers

• Semantic Topology: Defines knowledge as a sparse, editable, human‑readable conceptual graph

• Activation Dynamics: Introduces a meta‑reasoning model that operates through activation, inhibition, and resonance

• Five Core Primitives: Activation Arrays, Causal Signatures, Selection Pressure, Transform Learning, Semantic Abacus

• Transparent Reasoning: Enables full reasoning traces, error provenance, and verifiable outputs

• Modular Correctability: Allows knowledge updates without retraining and reasoning improvements without altering knowledge

• Implementation Pathway: Provides a clear roadmap from conceptual architecture to proof‑of‑concept system

Research Impact

This work contributes to AI architecture, cognitive science, epistemology, and human–AI collaboration by:

• Proposing a structure‑centric alternative to token‑based intelligence

• Offering a transparent, inspectable model of reasoning

• Introducing a modular architecture that separates knowledge from reasoning

• Providing a foundation for interpretable, evolvable AI systems

• Demonstrating how design cognition can generate novel AI architectures

• Establishing a multi‑paper research program for future development

Access and Documentation

ORCID: https://orcid.org/0009-0003-4876-9273

GitHub: https://github.com/Neuron-Soul-AI/Neuron-Soul-AI

Academia.edu: https://independent.academia.edu/MarceloTeixeira214

LinkedIn: https://www.linkedin.com/in/marcelo-emanuel-paradela-teixeira-702082382/

Email: marcelo.soul.ai@gmail.com

License: CC BY-NC 4.0

© Marcelo Emanuel Paradela Teixeira 2026