Open Cognitive Graph

Cognitive Infrastructure for Knowledge-Intensive AI

A research and collaboration handbook

Open Cognitive Graph (OCG) · AI Edtech Governance Trust (AEGT) · New Zealand opencognitivegraph.org · aegt.org

Preface: Why This Exists

Most AI systems today are built the same way: take a foundation model, connect it to a knowledge base, wrap it in a product. This works well for simple tasks. It breaks down in complex ones.

The failure is not the model. It is the absence of cognitive structure beneath the model — no explicit reasoning path, no evidence boundary, no domain-grounded decision logic. The model improvises. At small scale, improvisation is acceptable. At institutional scale, in high-stakes domains, it is not.

OCG exists to build what is missing: a layer of computable cognitive structure that lets AI systems reason verifiably, not just generate fluently.

This handbook explains what that means, how it is structured, what has been built, and how collaboration works.

Part I — The Problem

1.1 The Generation Gap

Large language models are remarkable at generating text that is fluent, contextually coherent, and often correct. They are significantly worse at:

• maintaining consistent reasoning across a long task
• staying within evidence boundaries set by domain standards
• producing outputs that are auditable against explicit criteria
• scaling the same logical structure across thousands of users without drift

These are not model capability problems. They are architectural problems. A model without structured domain knowledge beneath it is working without a skeleton. It can move, but not reliably, and not at load.

1.2 What the Real Difficulty Is

In complex knowledge domains — education assessment, environmental governance, professional knowledge systems, institutional decision-making — the hard problem is not whether AI can produce a plausible answer.

The hard problem is:

• Is the answer grounded in the right knowledge structure?
• Does it follow a reasoning path that experts would recognise as valid?
• Can you trace why the system produced this output rather than another?
• Will the same system produce consistent outputs across different users, contexts, and times?
• When the domain knowledge updates, does the system update cleanly?

None of these questions are answerable from generation quality alone. They require structure underneath generation.

1.3 Where Current Approaches Stall

Prompt engineering produces context-dependent behaviour. It does not produce domain-structured reasoning. It degrades across edge cases.

RAG (retrieval-augmented generation) improves factual grounding. It does not structure the reasoning path or constrain evidence boundaries. It retrieves; it does not reason.

Fine-tuning encodes domain knowledge into model weights. It makes knowledge opaque and difficult to update. It does not produce inspectable reasoning.

Ontologies and knowledge graphs structure knowledge well. They typically lack a mechanism for driving AI generation, routing, and decision-making from that structure.

OCG works at the intersection of these approaches — using structured knowledge to govern AI reasoning, rather than treating them as separate concerns.

Part II — What OCG Is

2.1 One-Sentence Definition

OCG is cognitive infrastructure: a framework for making expert knowledge, reasoning paths, evaluation criteria, and evidence boundaries computable — so AI systems can reason from structure, not improvise from probability.

2.2 The Core Mechanism: Cognitive Transition

The foundational unit of OCG is the cognitive transition:

c → r@D → C′

Where:

• c is the current cognitive state (what is understood, believed, or assessed to be true)
• r is a named cognitive operation (identify, reframe, bound, validate, classify, escalate...)
• @D is the task domain in which the operation is defined
• C′ is the updated cognitive state after the operation

This triple is not metaphorical. It is a computable structure. Every domain model in OCG is a graph of these transitions, with evidence conditions, boundary constraints, and outcome specifications attached.

The result is a reasoning path that is explicit, traceable, domain-grounded, and machine-executable — not a sequence of token predictions.

2.3 What "Computable Cognitive Structure" Means in Practice

When OCG structures a domain, it produces:

A knowledge ontology — the concepts, relationships, and hierarchies that define what exists in the domain.

A task domain model — the operations that can be performed in this domain, the conditions under which each applies, and the outputs each produces.

An evidence ledger — the empirical claims, benchmarks, and standards that bound what the system is licensed to assert. Every claim is tagged with confidence level and update conditions.

A decision graph — the routing logic that determines which operation to apply, which model or tool to invoke, and when to escalate to human judgment.

A governance action layer — the connection from AI output to institutional action: what the system recommends, why, under what conditions, and how outcomes feed back into the model.

Together, these constitute what we call a cognitive operating layer for a domain.

2.4 What OCG Is Not

OCG is not:

• A foundation model or model training system
• A RAG implementation or vector database
• A vertical AI application
• A content generation tool
• A benchmark or evaluation leaderboard

OCG is the structural layer beneath all of these. It can connect to any model, any retrieval system, any application. What it provides is the cognitive architecture that makes those connections principled rather than ad hoc.

Part III — The Three Layers

OCG is structured in three operational layers. Each layer represents a different depth of cognitive structure, and a different set of capabilities and use cases.

Layer 1 — Cognitive Structure

The knowledge and reasoning architecture

What it does: Structures domain knowledge as a navigable, computable graph of concepts, operations, and transitions. Makes the reasoning path from foundational knowledge to advanced understanding explicit and machine-readable.

The question it answers: What does it mean to understand something in this domain? What is the path from here to there? How do concepts relate?

Flagship instantiation: OCG Book — Cognitive Atlas

The Cognitive Atlas structures an entire knowledge domain as a cognitive transition graph. Rather than presenting knowledge as a curriculum (a list of topics in order), it presents it as a map of transitions: each node is a cognitive state, each edge is a named operation that moves understanding from one state to another.

The current implementation covers the path from school mathematics to AI literacy across nine conceptual chapters, 43 cognitive transition cards, and 13 distinct cognitive operation types. Chapter-to-chapter connections are explicit bridge structures identifying where one conceptual framework hits its limit and the next begins.

Also in this layer: AI Atlas — a full-stack map of the AI technology landscape, structured as a layered system from compute infrastructure (L1) to cognitive governance (L9). Each layer has defined concepts, failure modes, guardrails, and human roles. The cognitive transition graph exposes common misconceptions about where different kinds of AI work actually occur.

Who this layer is for:

• Research teams building AI literacy or cognitive development frameworks
• Institutions that need a structured representation of domain knowledge for AI systems
• Educators and curriculum designers working in AI-adjacent fields
• Anyone building systems where the knowledge architecture needs to be explicit and inspectable

Layer 2 — Cognitive Decision

The task routing and judgment architecture

What it does: Translates structured domain knowledge into operational decision logic. Given a task, determines what operation to apply, which tool or model to invoke, under what evidence conditions, within what boundary constraints.

The question it answers: For this task, in this domain, with this evidence — what is the right move, and why?

Flagship instantiation: OCG Bench — Task-Domain Decision System

OCG Bench is a decision graph system for AI model selection and task routing in complex professional tasks. It is not a benchmark leaderboard. It is a domain-structured routing engine.

For each task domain (single-file debugging, large-scale code migration, legal clause comparison, financial table extraction, literature review, and others), OCG Bench maintains:

• A cognitive transition graph specific to that domain
• Evidence-bound assessments of which models perform which operations acceptably
• Boundary conditions that determine when escalation is warranted
• An evidence ledger linking every routing decision to specific benchmarks, claims, and update triggers

The output of a Bench query is not "Model X scored Y on benchmark Z." It is: "For this task type, under these conditions, this model performs this operation acceptably — and here is the evidence. When these boundary conditions are exceeded, escalate to this alternative."

This is a structured decision, not a ranking.

Who this layer is for:

• AI engineering teams making model selection decisions in production
• Research teams studying AI capability boundaries and task-specific performance
• Platform teams building AI orchestration infrastructure
• Anyone whose AI deployment decisions need to be defensible, not just intuitive

Layer 3 — Cognitive Governance

The institutional action and accountability architecture

What it does: Extends cognitive structure all the way to institutional action — connecting AI diagnosis and recommendation to real operational decisions, with evidence chains, accountability mechanisms, and outcome feedback loops.

The question it answers: From assessment to action to verification: how does structured AI reasoning produce decisions that institutions can act on, defend, and learn from?

Flagship instantiation: OCG Assess — Regional Education Assessment Operations

OCG Assess structures the complete operational chain for higher-order cognitive assessment in education:

1. Evidence collection — assessment data, process records, qualitative indicators 2. CDC encoding — Cognitive Diagnosis Cards, each representing a specific pattern of cognitive difficulty, with confidence scores and evidence requirements 3. Card activation — when evidence meets threshold conditions, the relevant CDC activates 4. Action generation — the system generates specific institutional action recommendations (targeted teaching research, reassessment scheduling, escalation to regional tracking) 5. Outcome verification — results feed back into the graph, validating or updating the diagnostic logic

The system produces outputs structured for three different audiences simultaneously:

• For the regional education bureau: a governance dashboard — which schools, which patterns, which actions, which outcomes
• For the teaching researcher: a school capability portrait — structural weaknesses by cognitive dimension, with evidence
• For the domain researcher: a diagnostic card audit trail — why each card activated, what evidence supported it, whether the boundary conditions were met

Second instantiation: AIR Governance — Atmospheric Impact Assessment Ontology

A jurisdiction-portable ontology for atmospheric environmental impact governance. The core ontology covers the universal causal chain: project activity → emission source → pollutant → environmental pathway → receptor → assessment judgment → governance response.

Jurisdiction overlays connect this universal structure to specific regulatory frameworks: China (HJ 2.2, GB 3095), United States (Clean Air Act, NAAQS, NSR/PSD), European Union (IED, BAT/BREF), New Zealand (RMA, NES for Air Quality).

The same ontological structure — the same reasoning skeleton — governs AI-assisted assessment across regulatory contexts. Only the overlay changes.

Who this layer is for:

• Institutions that need AI outputs to drive real operational decisions, not just inform them
• Research teams studying AI governance, accountability, and institutional AI integration
• Policy bodies developing frameworks for AI use in regulated domains
• Education systems, environmental agencies, compliance functions, and professional services where AI recommendations carry institutional weight

Part IV — How the Layers Connect

The three layers are not separate products. They are a single architecture at three depths.

Layer 1 (Cognitive Structure) defines what exists in a domain and how it relates. It is the ontology and the knowledge graph.

Layer 2 (Cognitive Decision) uses that structure to govern task routing and model invocation. It is the decision engine running on the ontology.

Layer 3 (Cognitive Governance) extends decision logic to institutional action, with evidence chains, accountability records, and outcome feedback. It is the governance layer running on the decision engine.

A domain implemented at Layer 3 necessarily contains all of Layer 1 and Layer 2. A domain at Layer 1 can be extended to Layer 2 and Layer 3 as the application matures.

The mathematics remains constant across all three layers:

c → r@D → C′

The domain, the scale, and the institutional context change. The cognitive transition structure does not.

Part V — The Organisation

5.1 AI Edtech Governance Trust (AEGT)

OCG is a primary initiative of the AI Edtech Governance Trust (AEGT), an independent research trust incorporated in New Zealand (NZBN 50255431).

AEGT's mission is to strengthen human cognitive agency, support trustworthy AI-mediated systems, and develop the foundations of human-centred cognitive infrastructure for the AI era.

AEGT's work spans three methodological frameworks:

Open Cognitive Graphs (OCG) — Knowledge and semantic infrastructure: structured knowledge representation, semantic interoperability, explainable reasoning.

Endogenous Cognitive Topology (ECT) — Human agency and cognitive development: metacognitive training, judgment formation, cognitive resilience.

Metabolic Intelligence (MI) — Adaptive learning and collective intelligence: continuous learning, collective coordination, adaptive intelligence.

OCG is the computational infrastructure layer within this broader research programme.

5.2 Research Affiliations

OCG's development has been conducted in active collaboration with educational research institutions in China, including work affiliated with Ministry of Education engineering research centres. This collaboration grounds OCG's education instantiations in real institutional contexts and operational requirements, not only theoretical frameworks.

5.3 Current Status

OCG is in active research and prototype development. The instantiations described in this handbook — OCG Book, AI Atlas, OCG Bench, OCG Assess, and the AIR Governance ontology — represent working prototypes demonstrating the framework across different domains and institutional contexts.

The infrastructure is operational. Broad deployment awaits collaborative partnerships that bring domain expertise, institutional context, and research questions.

Part VI — Collaboration

6.1 What Collaboration Means

OCG is not a commercial software product seeking customers. It is a research infrastructure seeking co-developers.

The analogy that fits best: OCG is closer to a power grid than to an appliance. We build the infrastructure layer. What gets built on top — the applications, the research outputs, the institutional deployments — emerges from collaboration with partners who bring domain expertise, research problems, and institutional contexts we do not have.

Collaboration with OCG is joint development or joint research. You bring the domain. We bring the cognitive infrastructure. The outcome is something neither party would produce alone.

6.2 Who We Work With

Universities and research institutes working on education assessment, cognitive science, learning science, AI governance, AI literacy, or domain-specific AI applications. We provide technical infrastructure and methodological framework; you provide research questions, domain knowledge, and institutional access. Outcomes can include joint publications, shared datasets, demonstrated systems, and grant applications with OCG as the technical foundation.

Education technology teams building AI-powered learning pathways, adaptive assessment, ability diagnosis, personalised instruction, or cognitive development tools. OCG provides the cognitive architecture layer that makes outputs structured, verifiable, and scalable — rather than dependent on model improvisation at each inference.

AI research and engineering teams studying model capability, task-specific performance, AI reliability in production, or the boundaries of current AI approaches in professional domains. OCG Bench provides a structured framework for this kind of investigation.

Institutions in regulated or high-stakes domains — education governance, environmental assessment, compliance, professional knowledge systems — where AI outputs need to be explainable, auditable, and bounded by institutional rules. OCG's governance layer is built for exactly this requirement.

Policy and governance bodies developing frameworks for responsible AI use in institutional contexts. AEGT's governance research programme connects directly to the institutional architecture questions that effective AI governance requires.

6.3 What Collaboration Can Produce

Depending on the partnership, collaboration with OCG can produce:

• A joint research publication grounded in a working system
• A cognitive domain model for a specific field, usable as AI infrastructure
• A demonstrated operational system for a specific institutional context
• A grant application with OCG infrastructure as the technical foundation
• A pilot deployment generating evidence for broader rollout
• A shared dataset of structured cognitive transitions in a domain

6.4 How to Start

Most collaborations begin with a single concrete domain: one assessment context, one knowledge area, one task type, one institutional problem.

From that starting point, the cognitive model is built, the system is demonstrated, and the outcomes determine where to go next. We do not require a large commitment upfront. We require a concrete problem and a genuine interest in building something that holds up in practice.

Part VII — Expression for Different Contexts

OCG operates at a layer that sits between technical infrastructure and institutional application. Different audiences encounter it differently. The following are not scripts — they are accurate descriptions at different levels of technical depth.

For technical and infrastructure experts

OCG performs computable cognitive compilation. We structure tacit domain knowledge, task paths, and decision boundaries as a formal graph of cognitive transitions (c → r@D → C′), providing AI systems with an external cognitive operating layer that addresses long-range consistency and verifiability in complex task execution. The architecture is ontology-driven, evidence-bound, and domain-portable.

For domain researchers and institutional co-developers

Large language models have powerful language capabilities but no domain cognitive skeleton. OCG provides a cognitive structuring foundation. Where you have deep theoretical frameworks and evaluation standards — how to assess higher-order thinking, how to structure an environmental impact judgment, how a legal clause analysis should proceed — we have a method for compiling those frameworks into AI infrastructure. Together, we turn paper-based research theory into computable, scalable AI operational systems.

For institutional decision-makers and governance bodies

Current AI systems give you outputs. OCG-grounded systems give you complete evidence chains. The system not only produces a judgment — it shows why, against what criteria, under what boundary conditions, and what institutional action follows. Every output is traceable. Every recommendation is revisable. The system learns from outcomes.

Appendix — Current Instantiations at a Glance

Appendix — Key Terms

Cognitive transition — The fundamental unit of OCG: a structured movement from one cognitive state to another, via a named operation, within a specified domain. Formally: c → r@D → C′.

Cognitive operation (r) — A named, domain-specific action that transforms cognitive state. Examples: identify, reframe, bound, validate, classify, escalate, mechanize, locate.

Task domain (@D) — The domain context in which an operation is defined. Operations have different meanings and constraints in different domains.

Evidence ledger — The structured set of empirical claims, benchmarks, and standards that bound what the system is licensed to assert. Every claim has a confidence level and update conditions.

Cognitive Diagnosis Card (CDC) — In the assessment layer: a structured pattern representing a specific cognitive difficulty, with evidence requirements, confidence threshold, and associated action recommendations.

Cognitive operating layer — The complete set of structures — ontology, task domain model, evidence ledger, decision graph, governance action layer — that constitutes a domain implementation of OCG.

Domain overlay — A jurisdiction- or context-specific extension that applies to a universal ontological structure. Used in AIR Governance to adapt universal impact assessment logic to specific regulatory frameworks.

Open Cognitive Graph is a primary initiative of the AI Edtech Governance Trust. AEGT · NZBN 50255431 · Incorporated in New Zealand · aegt.org OCG · opencognitivegraph.org