Understanding is not recall, and a system is not its parts. The methods here are how we tell the difference — and build for it.

Premise

Most systems for learning and for managing knowledge rest on two assumptions so widespread they are invisible: that to know is to recall, and that a system is the sum of its components.

Neither holds. Understanding means being able to derive a thing from its foundations, explain it in one's own words, apply it where it has not been applied before, and teach it to someone else. And what makes a person, an organisation or a knowledge system healthy is not the state of its parts but the condition of the medium that connects them and the loops that regulate them.

We present the methodology because it explains why the systems we build behave as they do — not because clients need to become theorists. The research is rigorous; the implementation is practical. One does not need to understand pharmacology to benefit from medicine.

How understanding forms

Gordon Pask's Conversation Theory, developed over thirty years of experiment, supplies the core mechanism.

Understanding is not transmitted; it is constructed, in a particular kind of structured conversation in which the learner must explain, derive and demonstrate rather than receive and repeat. Pask showed experimentally that when a learner teaches back — explains what they have learned in their own terms and derives it from first principles — the understanding that results is deeper, more durable and more transferable than any form of content delivery produces. Knowledge is held as a mesh of concepts linked by what each requires to be understood, with many points of entry, rather than as a sequence.

In our systems, assessment asks a person to explain and teach, not to choose from a list. The same test is applied to the AI agents: before any of them writes to a graph or acts on a person's behalf, it must state back what it understands it is about to do. Teachback is both a learning principle and a safeguard.

How a system regulates itself

From Ross Ashby: a regulator must have as much variety as what it regulates, and every good regulator of a system must be a model of that system.

These two laws decide the shape of the architecture. Agents are organised in layers because no single agent can carry the variety of a person's or an organisation's knowledge; each layer regulates the one below it and is regulated by the one above. Provenance is recorded on everything because a regulator that cannot see what it governs cannot govern it. Mick Ashby's extension — the ethical regulator, and the law that ethical adequacy is never finally achieved — is why governance is a standing function of the system rather than a policy document, and why a person remains at the point where meaning is decided.

James Bryant's behavioural cybernetics carries Ashby's law into conduct: only behaviour can regulate behaviour, and every behaviour has ten dimensions — person, state, understanding, decision, action, way, time, position, resource, outcome — which must each be right for an outcome to follow. We use them to ask what has to be in place for a person, a team or an agent to behave as an outcome requires, and to design so that the right things are easy and the wrong things hard. Stafford Beer's viable system model gives the recursive form: the same organisation of operations, coordination, control, intelligence and policy at every scale.

How trust in AI is earned — record-keeping and standards

Trust in a system that uses AI is not asserted; it is evidenced. The evidence is a record of what was done, by what, on what authority, and to what standard.

Provenance in our architecture is therefore not only an epistemic principle — knowing how a thing came to be known — but a governance one. Every proposal an agent makes, every model call, every human approval and every change to a default is recorded, hashed and kept. Nothing is deleted; it is superseded. That record is what allows a person, an auditor or a regulator to ask what the system did and why, and to receive an answer rather than an assurance.

The external frameworks give that record its shape and let conformance be shown rather than claimed. They differ by jurisdiction, and a knowledge system that may serve a person in one country and an organisation in another must be able to hold all of them.

Internationally, ISO/IEC 42001:2023 defines an artificial intelligence management system — policy, roles, risk assessment, AI impact assessment, controls and continual improvement — on the same pattern as the quality and information-security standards organisations already run. It is the standard against which organisations are now certified; ISO/IEC 42006 sets the requirements for the bodies that certify them, and ISO/IEC 23894 gives the risk-management guidance the management system relies on.

In the United States, the NIST AI Risk Management Framework (AI RMF 1.0, 2023) is voluntary rather than certifiable. It organises the work of trustworthiness into four functions — govern, map, measure, manage — and names the properties a trustworthy system must show: valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias managed. Its 2024 generative-AI profile applies the same functions to the models our agents call.

In the United Kingdom, the approach is assurance and accredited certification rather than a single statute. BSI became, in January 2026, the first certification body accredited by UKAS to certify organisations against ISO/IEC 42001. The Alan Turing Institute's Trustworthy and Ethical Assurance method supplies the form of the argument: a structured assurance case in which a claim about a system's properties is supported by evidence, so that the case can be read, challenged and revised rather than taken on trust. Together they make record-keeping the substance of assurance, not an afterthought to it.

In the European Union, the AI Act (Regulation (EU) 2024/1689) is binding law with a risk-based structure. Its prohibitions have applied since February 2025 and its obligations on general-purpose AI models since August 2025; transparency duties for AI-generated content apply from August 2026. Under the Digital Omnibus agreed in May 2026, the obligations on stand-alone high-risk systems now apply from December 2027 and those on AI in regulated products from August 2028. For high-risk systems the Act requires a risk-management system, data governance, technical documentation, automatic record-keeping and logging, transparency to users, human oversight, and accuracy, robustness and cybersecurity — with harmonised European standards, drafted through CEN-CENELEC, giving a presumption of conformity. ETSI's work on AI supplies further technical standards where they apply.

In the architecture these frameworks are held as first-class objects — standards, policies, risks and controls in the graph, linked to the agents, capabilities and projects they govern — so that a governance function can produce a conformance view rather than an opinion. This is Ashby's ethical regulator made operational: the standard is the model the regulator must contain, and the record is how it sees what it governs.

Where the observer belongs

Second-order cybernetics, in the line from Heinz von Foerster to Ranulph Glanville, places the observer inside the system observed.

Glanville's radical version holds that a black box is a construct of its observer, that every whitening of a box yields two more, and that knowing is therefore always a conversation between observers rather than a view from nowhere. Design, on this account, is a conversation with a situation and with oneself — not the solving of a given problem. Ethics follows from the recognition of not-knowing: one acts so as to keep the conversation open.

This is why the architecture is a society of agents rather than a pipeline; why a supervisory function observes the agents that do the work; why the system keeps a record of its own building; and why nothing in the graph is deleted, only superseded — so the system can observe its own history. It is also the standard by which we judge whether a collaboration, a meeting or a document is genuinely conversational or merely transmits.

How meaning is read and new action found

Charles Sanders Peirce supplies the logic of signs and the only logical account of how genuinely new hypotheses are generated.

An issue, a conversation or a document is treated not as a brute fact to be managed but as a sign pointing beyond itself — to be read through Peirce's three categories of possibility, actuality and law-or-habit, and interpreted for what it demands. Abduction, the inference to the best explanation, is how our reasoning agents generate candidate connections, hypotheses and courses of action rather than merely retrieving what is already there. Pragmatism supplies the test: the meaning of a claim is the difference it would make to conduct, and a course of action counts only if it would become a new habit.

How a system anticipates

Robert Rosen's anticipatory systems theory and relational biology explain why a living system acts on a model of its future, not only on its present.

An anticipatory system contains a model of itself and its environment and changes its present state in light of what the model predicts. Rosen's metabolism–repair framework distinguishes threats to a system's core function from threats to its capacity to maintain itself, and shows why complex systems cannot be captured by simulation alone. We use this to ask whether a person's or an organisation's internal model is adequate to the complexity it faces, whether an issue threatens what it does or its ability to keep doing it, and where the model must be enriched — which is what a knowledge graph with provenance and time is for.

How reason and process are kept alive

Alfred North Whitehead supplies both an ontology — reality as process, not substance — and a diagnosis of how reason works or fails.

In Whitehead's process view, what exists are occasions of experience, internally related, each taking up its past and adding something new. Much systems language borrows process words while keeping substance thinking; we examine documents and designs for whether they embody process — internal relation, becoming, participation — or merely name it, and for the fallacy of misplaced concreteness that treats abstractions as things.

In The Function of Reason, reason exists to promote the art of life: to live, to live well, to live better. It operates in two modes — the practical cunning that gets the next thing done and the speculative vision that asks what for — and advances when the two meet. Methods fatigue; obscurantism defends them; purpose gets excluded. We use this to judge whether a strategy, a policy or a knowledge system is still reasoning, or merely maintaining itself.

How wholes are seen — going upstream

Henri Bortoft, drawing on Goethe's way of science and on phenomenology, distinguishes the finished product of understanding from the living act in which it comes into being — and shows that we habitually work only with the former.

Downstream is where knowledge arrives as results: definitions, categories, documents, the counterfeit whole assembled by adding parts together. Upstream is the dynamic in which meaning is still appearing — the whole presencing in each part, the part as a place where the whole shows itself. Bortoft's claim is that the whole cannot be reached by summing parts, nor by standing back from them; it is met by dwelling in the parts with a different quality of attention, so that the whole comes to presence through them. Understanding is then an event, not an object.

This is why our systems record how a thing came to be known and not only what is known: provenance is the upstream of every fact. It is why the unit of record is the conversation rather than the extracted claim, why nothing is deleted but only superseded, and why a knowledge graph is treated as a place where a whole way of thinking can presence, rather than a store of its results. It is also the sensibility behind Ground Regulation: attend to the medium in which things come into being, not only to the things.

How the ground is kept healthy

Ground Regulation, derived from Alfred Pischinger and Hartmut Heine's biology of the extracellular matrix, holds that dysfunction originates in the connective medium, not in the components.

Cells live in a ground; so do people and organisations. When the ground degrades, the components fail, and treating the components — the Virchow error — rearranges the failure without reaching its cause. The framework assesses nine dimensions of ground condition across four stages of degradation and applies them isomorphically to the individual and the organisation: the same dimensions, in parallel, sharing one ground. We use it to ask whether a plan, a review or a design starts from the ground or from abstraction, and whether its interventions would restore ground conditions or merely move structure about. It is also the sense in which we say that knowledge is relational: the connections are the ground, and the graph exists to keep them well. The inquiry that drives the practice runs the same span: how living entities and the matrices they exist in communicate, regulate, perturb, resonate, adapt and transform across scales — from cell and extracellular matrix, through human consciousness and relationship, to community, organisation and governance.

How capability develops, and how to see it

Elliott Jaques describes how a person's capacity to handle complexity grows through identifiable stages; Gillian Stamp shows how to recognise where a person stands without reducing them to a score.

People at different stages process information at different levels of abstraction, over different time horizons and across different systemic scope. Expert knowledge is characteristically complex, layered and systemic, so a system that does not account for where a learner stands either overwhelms or bores them. Jaques's framework lets material be calibrated to the person and stretched from there. Stamp's appreciative method — seeking what a person brings rather than what they lack — lets a system notice, through ordinary conversation, how each person works and how to support the next step. Conversation yields more reliable evidence of capability than any multiple-choice instrument, and the two frameworks apply directly to leadership development and succession.

How people, communities and action are read

Knowledge lives in people, in the communities they form, and in what they do. Three complementary analyses read a situation for each.

Stakeholder analysis asks who is involved, what they bring, what they need, and how power and relationship run between them. Étienne Wenger's communities of practice asks how a community actually functions — its shared domain, its patterns of participation from core to periphery, its repertoire of ways of working — and whether newcomers can legitimately join at the edge. The action framework asks what is happening now, what is committed, how work habitually gets done, and what is missing. Applied to a transcript, a meeting or a plan, the three together show what the situation contains, what needs learning, and what has to happen next.

Where a group must think together, we select the structured dialogue method the situation requires: Beer's Syntegration for a whole system deliberating on its own future; Warfield and Christakis's Structured Democratic Dialogue for surfacing and structuring the interdependencies among a group's own observations; Harrison Owen's Open Space where the agenda must emerge from the participants; Juanita Brown and David Isaacs's World Café where what is needed is cross-pollination of many small conversations around questions that matter. The choice is itself a design decision, made from the situation rather than from habit.

Together

These strands are not a reading list. They are distributed through the architecture as its working lenses, and what carries them is an organisation of AI agents rather than a single model. Each agent holds one role and is accountable for one kind of work — capturing a conversation faithfully, placing it in context, proposing what it meant, logging where each proposal came from, reasoning over what is now known, noticing what has changed, checking that the right things were done in the right order. No agent writes to the graph on its own authority; each proposes, states back what it understands, and hands on. They are arranged in three layers that answer three questions — what happened, what does it mean, how is the system changing — and each layer learns from the one below it, so that what the reasoning layer finds improves how the operational layer works, and what the meta-learning layer notices improves both. Above the operational work a supervisory agent observes the others, as second-order cybernetics requires; a person sits at the one point where a proposal becomes a fact.

The lenses are what the agents think with. Conversation Theory governs every dialogue and every act of writing to the graph. Stakeholder, community and action analysis read what a conversation contains. An ensemble of reasoning lenses — Pask, Ashby, Rosen, Whitehead, Peirce — interprets what it means, generates hypotheses, and tests whether the system's model is adequate to its situation. Second-order cybernetics, Bortoft's upstream attention, Ground Regulation and the function of reason watch the system itself: whether it includes its observer, whether it stays with the coming-into-being of meaning or only its results, whether its ground is sound, whether it is still reasoning. Ashby's ethical regulator and Bryant's ten dimensions shape governance and conduct, and ISO/IEC 42001, the NIST AI Risk Management Framework, the Turing Institute's assurance method and the EU AI Act give the record its external form. Jaques and Stamp calibrate the system to the person.

Which AI performs which role is a separate matter. A registry maps each capability the organisation of agents needs to whichever model or tool currently supplies it best — a cloud model, a local model on the owner's own hardware, a specialist skill — and routes private material only where it may go. The models are implementations; the organisation of agents, and the lenses it thinks with, are the architecture.

The technology to hold this together — conversational AI, graph databases, capable local models — has only recently become possible. The theory has been ready for decades. The methods are refined in use, and this page will change as the work does.

Reading

• Conversation, Cognition and Learning — Gordon Pask, 1975

  Conversation Theory in education: teachback, entailment structures, learning strategies.

• An Introduction to Cybernetics — W. Ross Ashby, 1956

  Variety, regulation, and the requirements on any system that would govern another.

• Ethical Regulators and Super-Ethical Systems — Mick Ashby, 2020

  The ethical regulator theorem and the law of inevitable ethical inadequacy.

• The Black Box; Try Again. Fail Again. Fail Better — Ranulph Glanville, 2009 (collected papers)

  Second-order cybernetics, the black box as an observer's construct, design as conversation.

• Collected Papers — Charles Sanders Peirce

  Semiotics, the three categories, abduction and pragmatism.

• Anticipatory Systems — Robert Rosen, 1985

  Systems that act on models of their future; metabolism and repair.

• The Function of Reason — Alfred North Whitehead, 1929

  Reason as the promotion of the art of life; practical and speculative reason; the fatigue of method.

• Process and Reality — Alfred North Whitehead, 1929

  The process ontology: actual occasions, prehension, internal relations.

• The Wholeness of Nature — Henri Bortoft, 1996 · Taking Appearance Seriously — Henri Bortoft, 2012

  Goethe's way of science; the authentic and counterfeit whole; upstream and downstream thinking.

• Matrix and Matrix Regulation — Alfred Pischinger, ed. Hartmut Heine

  The biology of the extracellular matrix on which Ground Regulation rests.

• ISO/IEC 42001:2023 — ISO/IEC · AI Risk Management Framework 1.0 — NIST, 2023 · Trustworthy and Ethical Assurance — Alan Turing Institute · Regulation (EU) 2024/1689, the AI Act — European Union

  The international management-system standard, the US risk framework, the UK assurance-case method, and the EU statute that give AI record-keeping its shape.

• Requisite Organization — Elliott Jaques, 1989 · Human Capability — Jaques and Cason, 1994

  The development of capacity for complexity, and its alignment with organisational structure.

• The Tripod of Work — Gillian Stamp, BIOSS

  Appreciative assessment of how people handle complexity at work.

• Beyond Dispute — Stafford Beer, 1994 · A Science of Generic Design — John Warfield, 1994 · Open Space Technology — Harrison Owen, 1997 · The World Café — Juanita Brown with David Isaacs, 2005

  Syntegration and the viable system; structured democratic dialogue; open space; conversations that matter.

• Communities of Practice — Étienne Wenger, 1998

  Social learning in communities: domain, community, practice.

Contact

To discuss how these methods apply to your work, write directly. We respond to every enquiry, and every engagement is held in confidence.

peter.tuddenham@coexplorer.com