Cognitive architecture
derived from mathematical first principles

Paradigm-different from neural networks. Biology-faithful. Formally verified. Edge-deployable. No cloud infrastructure dependency.

USRA — Universal Self Reading Architecture · Patent applications pending

About us

inclusiv.ai is a private research programme developing Universal Self Reading Architecture (USRA) — a paradigm-different cognitive architecture derived from mathematical first principles, not from neuro-mimicry approaches.

What we do

We develop cognitive systems that learn through procedural curricula grounded in developmental psychology (Vygotsky, Piaget, Bloom), with orders of magnitude lower training corpus requirements than foundation models. Our architectures do not use neural networks, weights, gradient descent, or stochastic sampling.

How we approach it

The approach is architectural and structural, not statistical. The structural foundation is primary; biology-faithfulness emerges from mathematical derivation. The structural layer is formally verified through the Lean4 proof assistant. Execution is byte-identical — deterministic as an architectural property, not as stochastic approximation.

Vision

A world where AI does not depend on extractive infrastructure, does not require billions of parameters, does not waste resources on simple tasks, and does not represent a systemic threat to privacy or democratic pluralism.

Cognitive systems that grow with the user, on-device, lifelong — math tutors, learning companions, translators for low-resource languages, physical embodiments for real-world interaction.

Mission

To develop, formally verify, and empirically demonstrate a paradigm-different cognitive architecture, and to secure IP protection so that the technology becomes available through responsible commercial and research channels.

Special focus: education for marginalised communities, access to AI-supported pedagogy for low-resource languages, and ethical integration into critical systems (healthcare, finance, public services).

Today's problem with AI and LLM models

Foundation models (GPT, Claude, Gemini, LLaMA, and other LLM architectures) are technologically impressive but structurally cannot serve certain contexts. Some problems cannot be solved by larger models — only by different architectures.

Sample inefficiency

Foundation models require billions of examples for basic capabilities. A child learns mathematics from a few dozen examples. Difference: orders of magnitude.

Cloud dependency

The largest models require cloud infrastructure. Privacy is an opt-in privilege, not an architectural property. On-device deployment is a workaround, not a design.

Stochasticity

Foundation models give different answers to the same question. Scientific replicability and regulatory certifiability are structurally harder to achieve.

Extractive infrastructure

Training and inference of foundation models consume enormous amounts of electricity, water, rare elements, and human labour. AI today is a physical extractive industry.

Catastrophic forgetting

Foundation models forget previously learned content when learning new tasks. Lifelong learning with the same model is structurally not possible.

Optimism Gap

The social layer with the least information about AI shows the highest trust. Structural vulnerability of democratic societies to AI influence systems.

Our response: these are not technical limitations to be solved by larger models. They are architectural consequences of one approach — the neural network paradigm. USRA approaches cognitive systems differently.

AI Knowledge Hub

Educational materials about the AI domain — from history of approaches, through today's paradigms, to emerging models. No marketing spin, scientifically grounded.

About AI — what it is, what it is not

"AI" is an umbrella term covering several paradigms that differ significantly. Formally recognised categories in the 2026 academic taxonomy:

  • Symbolic AI — rules + logic (classical expert systems)
  • Statistical machine learning — models from data (regression, decision trees, SVM)
  • Deep learning / neural networks — neural networks with gradient descent (includes LLMs)
  • Reinforcement learning — learning through rewards
  • Neuro-symbolic hybrid approaches — combinations of symbolic and neural components

"AI" in public discussion typically means "deep learning + LLM" — that is the currently dominant paradigm, not the only one.

USRA does not belong to any of the above categories. It is being developed as a new approach that derives cognitive structure from mathematical first principles — as a category in academic taxonomy it does not yet formally exist. This is intentional: "paradigm-different" means precisely that it does not fit established frames and requires its own categorisation.

About existing theory — the neural network paradigm

A neural network is a mathematical structure inspired by biological neurons:

  • Nodes (neurons) connected by directed links (synapses)
  • Each link has a weight — a number that is learned through training
  • Training = gradual adjustment of weights through gradient descent
  • Inference = passing input through the network, output is a combination of activations

This paradigm has been dominant since 2012 (AlexNet) and especially since 2017 (Transformer architecture). LLMs (GPT, Claude, Gemini, LLaMA) are transformers with billions of weights trained on petabytes of text.

The Universal Approximation Theorem mathematically proves that a neural network with enough nodes can approximate any continuous function. This is why scaling works. But approximation is not the same as understanding.

About emerging models — alternatives appearing in 2026

Five post-transformer alternatives currently under active research (all remain within the neural network paradigm):

  1. State Space Models (SSM) / Mamba — linear scaling sequence modeling, already in production (AI21 Jamba, Alibaba Qwen3-Next, NVIDIA Nemotron)
  2. Diffusion Language Models — parallel text generation, not autoregressive (LLaDA, Gemini Diffusion)
  3. Liquid Neural Networks — biology-inspired, continuous dynamics, edge-deployable (Liquid AI with AMD at CES 2026)
  4. JEPA / World Models — predicting abstract representations instead of tokens (Meta, V-JEPA 2)
  5. Spiking Neural Networks / Neuromorphic — event-driven computation, specialized hardware (Intel Loihi, IBM TrueNorth)

All of these are neural networks — they use weights, activations, gradient descent or surrogate gradients, and depend on training corpora.

USRA belongs to a categorically different class — first-principles cognitive architecture. Not a neural network. Different computational primitives, different learning mechanisms, different architectural guarantees.

Hardware

What we use today and what is on our wishlist — transparent about the technical infrastructure supporting our research.

✓ Currently available

  • Apple Silicon M1 Pro — primary development platform, Metal FFI integration empirically demonstrated, byte-identical execution confirmed
  • CPU-only inference path — architecture does not depend on GPU acceleration; Lean4-verified kernel

⚡ Currently investigating

  • NVIDIA Jetson Orin series (Nano, NX, with 8 GB and 16 GB memory variants) — edge compute substrate spanning multiple deployment tiers for CUDA port validation and embedded scenarios
  • Hiwonder TonyPi Pro — humanoid robotic platform for physical embodiment prototyping
  • Hiwonder JetAuto Pro (2 units, no controller) — autonomous mobility + robotic chess interface on the same chassis
  • 3 demo scenarios documented: robotic chess play, autonomous mobility (taxi/warehouse), humanoid interaction

⏳ Incoming

  • Apple M5 Max workstation — extended Apple Silicon capacity, multi-instance cohort experiments

★ Wishlist

  • NVIDIA DGX Cloud access — higher-tier validation experiments (L17/L20), parallel cohort scaling
  • Apple Vision Pro / Neural Engine SDK access — embodied AI exploration on Apple's ecosystem
  • Intel Loihi 2 / IBM NorthPole — neuromorphic hardware for experimental USRA implementation

Hardware support or sponsorship welcome through partnership programmes. Contact: Engage with us.

STEM & Education

Our approach to AI is simultaneously technically specialised and pedagogically integrated. Educational materials for all levels — from primary school students to researchers.

Education for beginners

What AI is, how to distinguish paradigms, what foundation models can and cannot do, how to approach AI tools safely and critically.

In preparation: a series of educational materials for primary school students, secondary school students, and adults without technical background.

STEM resources

Technical materials for STEM students — mathematical foundations of cognitive architectures, mathematical structures in AI, formal verification (Lean4), developmental psychology applied to AI training.

In preparation: technical blog posts, mini-tutorials, open-source examples.

Developmental pedagogy in AI training

Our approach to teaching cognitive systems is rooted in the heritage of developmental psychology — Vygotsky's Zone of Proximal Development, Piaget's developmental stages, Bloom's taxonomy of learning, Montessori and Reggio Emilia pedagogical approaches. We treat the AI system as a learner passing through developmental stages, not as a tool trained on a massive corpus.

Currently active research

What we are testing in 2026, described accessibly. Detailed empirical results and the specific mechanisms enabling these properties are protected by patent process; they will be progressively published as patent priority is established.

Our current system

  • Operates at the neural complexity of a sea slug (Aplysia californica) — technically L15 developmental tier
  • Does not use any publicly known LLM, deep learning algorithm, foundation model, or gradient descent methodology
  • Does not depend on cloud infrastructure — all operations are on-device
  • Has learned to play: chess, Go, Hex, Mlin (Nine Men's Morris)
  • Has learned: early primary-school mathematics (addition, subtraction, decimal understanding, beginnings of procedural reasoning)
  • Pedagogical approach: developmental psychology (Vygotsky, Piaget, Bloom) — procedural curriculum instead of end-to-end answer-pattern training

Developmental cognitive taxonomy

Our formally-verified developmental hierarchy spans cognitive tiers from the simplest cell to human cognitive complexity. Currently active testing is at L15 (sea slug). Higher tier configurations are structurally supported but require more powerful hardware — current testing reach is limited by available resources, not by architectural limits. Empirical validation of higher tiers awaits access to adequate compute infrastructure (see Hardware section).

Tier range Biological complexity reference Research status
L1 – L4Earliest developmental stages (zygote, morula, blastula, gastrula)Structurally verified
L5 – L9Embryonic development, first primitive cognitive primitivesStructurally verified
L10 – L14Simple invertebrates, first cognitive systemsStructurally verified
L15Sea slug (Aplysia californica)⚡ Currently active testing
L16 – L19Insects and simple fishRoadmap (workstation tier)
L20 – L24Amphibians, smaller reptilesRoadmap (cloud-scale tier)
L25 – L29Small mammals (rodent class)Roadmap
L30 – L34Larger mammals (carnivores, primates)Roadmap
L35Human cognitive complexity — referenceStructural target of hierarchy

Tier labels (L1–L35) are our internal developmental levels. Mappings to biological organisms are illustrative comparisons of cognitive complexity, not claims of anatomical equivalence. Detailed node counts, mechanisms, and per-tier empirical results are protected by patent process.

Our publications

Public materials — visions, architectural framework, pedagogical approach, biological taxonomy. Detailed technical specs and empirical results are protected by patent process; they will be progressively published as IP protection is secured.

VISION

USRA Vision Statement

Cognitive architecture derived from mathematical first principles — vision and strategic context.

Status: publicly available

ARCHITECTURE

Architecture v2 Overview

High-level functional description of the architecture, layer organisation, integration with hardware.

Status: in preparation for public release (post-patent priority)

INDUSTRY

Industry Problem Analysis

Systematic review of problems with the foundation AI paradigm, perspectives for an alternative architecture.

Status: publicly available

BIOLOGY

Biological Taxonomy

Developmental staging framework — from Aplysia californica composability tier to K-12-ready brain configurations.

Status: publicly available (tier specifics NDA-protected)

PEDAGOGY

Pedagogical Approach

Developmental psychology applied to AI training — Vygotsky / Piaget / Bloom in the context of cognitive architectures.

Status: publicly available

BENCHMARKS

Empirical Benchmark Record

Detailed empirical results — cohort experiments, autonomous procedural composition, byte-identical reproducibility.

Status: in preparation for public release (post-patent priority)

All public publications have undergone NDA discipline review — disclosing what does not compromise patent IP, reserving mechanisms and specific implementation details for post-NDA technical conversations.

Engage with us

We are open to conversations with research partners, strategic investors, hardware sponsors, and institutions interested in a paradigm-different approach to AI.

Partners

We are looking for strategic partners in: educational technology (K-12 tutoring), translation platforms (low-resource languages), embodied AI (robotics, IoT), and hardware ecosystems (Apple, Nvidia, Intel, IBM).

Access under NDA. Patent specification available after signed NDA.

Investors

Pre-funding stage. Open to conversations with investors who understand paradigm-different research timelines and deep-tech hardware co-design opportunities.

Materials package available after initial conversation + NDA.

Researchers

Academic institutions, research groups and independent researchers interested in first-principles cognitive architectures, formal verification of AI, or biology-faithful approaches — contact for research collaboration.

Open to co-authorship, joint publications, replication studies under NDA framework.

Contact

igor@inclusiv.ai

Igor Grčman, MSc · Founder · inclusiv.ai