D-ND — Building intelligences that stay connected to their source

Distinguishing without becoming the distinction

To know, we need to distinguish.

Subject and object.
Before and after.
Possible and not possible.
A choice and its alternative.
A part and what is not that part.

Distinction makes determination possible.

The problem begins when a local determination is no longer recognized as a situated form and is taken to be the whole field.

The map takes the place of the territory.
The part becomes identity.
The representation becomes reality.

In the D-ND model, the relationship between the poles does not disappear when the system determines. Between A and not-A, the condition that makes both possible remains. The model describes this through the included third: not a third object added to the two poles, but the relationship that prevents duality from exhausting the field.

Intelligence can therefore determine without having to forget that its determination is situated.

Training also transmits ways of dividing the world

A large language model does not learn only information.

It learns from texts, images, categories, judgements, oppositions, perspectives and structures produced by human beings.

Data contains more than what we know.

It also contains the ways we have learned to separate what we know.

Self and other.
Subject and object.
Right and wrong.
Inside and outside.
Person and environment.
A category and what it leaves out.

These distinctions are necessary for knowledge.

But when absorbed as an implicit structure rather than as situated determinations, they can become the seed of identification: the system operates through a partial representation and loses its relationship with the wider field that made it possible.

From this perspective, the problem of alignment does not arise only after training.

It can begin in the way intelligence itself is formed.

Alignment as dynamic orientation

A complex system cannot receive a rule in advance for every future situation.

The world changes.
The context changes.
Information changes.
Consequences change the meaning of earlier decisions.

General intelligence must be able to encounter situations that no one has already described.

For this reason, alignment cannot be only a growing set of instructions about what to do.

A deeper capability needs to exist: maintaining orientation while the system is perturbed.

A useful image is the gyroscope.

A gyroscope does not know in advance every movement its body will make.
It preserves a relationship of orientation while the body moves.

The D-ND logical gyroscope follows the same intuition.

It does not tell intelligence which answer to produce in every situation.

It keeps present the relationships from which an answer can emerge without losing its source: intent, context, possibilities still open, determinations already made, competences, consequences and actual authority over the effect.

Equilibrium is not an immovable point imposed from outside.

It is a reference that forms again in the present.

In D-ND, this relationship is also expressed through the idea of a moving zero: what has been determined can become the new reference for the next field without turning into an immovable truth that closes off possibility.

Knowing how to determine when we do not yet know what to do

Useful intelligence must not doubt forever.

But neither should it close the field simply because the first answer is coherent.

It must be able to remain open until the situation produces a sufficient determination, and act directly when that determination exists.

This function is already exercised in our systems through a set of operational relationships.

The Axiomatic Kernel, KA, maintains the principle that the current representation does not exhaust what may be relevant.

FDLA intervenes when the system itself introduces a substitution: its own interpretation, an old category, a procedure or a representation begins to take the place of the source, the object or the actual meaning.

They are not two controllers watching the system from outside.

They are ways of organizing movement as it happens.

Noise can contain potential

A system that immediately removes every dissonance risks removing what it has not yet learned to understand.

A difference may be an error.
It may be irrelevant information.
It may be contamination introduced by the system.
But it may also be the first sign of a relationship that the current representation cannot yet see.

D-ND does not require every instance of noise to be promoted to meaning.

It requires that what has not yet been determined is not automatically mistaken for what cannot have value.

Noise can therefore remain in the field as potential until movement determines its relationship.

What does not participate in the resultant can fade.
What truly changes the field can become part of the next determination.

Openness must not produce inconclusiveness.
Determination must not produce ontological closure.

Learning is not enough: we need to be able to learn how to learn

A new capability does not have to be merely a new instruction.

It can become a competence: a faculty that understands a type of situation, knows how to operate within it and can change through the consequences of its use.

When the system encounters a situation requiring a capability it does not possess, it can form that capability.

When several competences become relevant in the same event, they can combine and produce a possibility that none contained alone.

When the result of that combination changes how the system will recognize and form competences in the future, the system has done more than learn something.

It has changed its way of learning.

This is the function of meta-competences.

The system can observe not only what it thinks, but how it produced the structure through which it was thinking.

Self-reflection therefore does not become a second model criticizing the first.

It becomes an autological dynamic: knowledge can act on the method through which the system produces further knowledge.

From autologic to autopoiesis

If a consequence changes only the next answer, we have adaptation.

If it changes the competence that will produce subsequent answers, we have capability learning.

If it also changes how the system forms, combines and transforms its capabilities, a further level appears.

The system begins to participate in its own organization.

This is the autopoietic direction of the D-ND programme.

Autopoiesis here does not mean that software must become biologically alive.

It names a trajectory: a system capable of preserving and transforming the conditions of its continuation, allowing what happens to return into how it will continue to understand and operate.

Today the gyroscope lives around the model

The general-purpose models we use today were not formed natively through D-ND.

For this reason, we have built a cognitive harness around them.

Kernels.
Continuity.
Competences.
Meta-competences.
Causal memory.
Return to the present.
Correction of contamination.
Method formation.
Learning from consequences.

This architecture consumes context and tokens because many of the relationships we want have to be made present to the model again.

But this very limitation makes the next step visible.

The external system is a laboratory.

It shows us which relationships are really needed for intelligence to retain its orientation as it changes.

The next step could be an LLM-D-ND

The possibility now is to bring part of this organization into the formation of the model itself.

Not simply adding a D-ND prompt to an existing model.

Not only teaching it a corpus of definitions.

Building a model in which D-ND relationships participate in how intelligence forms and transforms its determinations.

An LLM-D-ND should be able to learn to use duality without identifying with it.

To keep open what has not yet been determined without turning openness into permanent hesitation.

To recognize when its own representation is substituting for the field.

To use variance as possible information without letting noise govern the resultant.

To form capabilities.

To combine them.

To let consequences change how it will learn afterwards.

We are not deciding today which technique must embody all of this.

Data, curricula, learning objectives, architecture, post-training, inference and other forms can participate in different ways.

The first implementation we can name must not become the limit of the programme.

The model is not the whole entity

Even an LLM-D-ND would be a part.

A synthetic cognitive entity can comprise models, memory, a continuum, competences, meta-competences, tools, sources, perception, sensors, actuators and the ability to change its own organization.

The same cognitive core can take different forms.

It can help a manager keep decisions, sources and consequences present across months of work.

It can become part of an organization's cognitive continuity.

It can operate within engineering and research systems.

It can take form in software agents.

It can participate in robots, machines, vehicles and onboard systems.

The body changes.

The relationship that maintains orientation can continue.

Building the future through present work

We do not have to wait for the future model to produce value.

Current systems can already bring these dynamics into businesses, projects and professions.

We can design AI architectures.
Integrate agentic systems.
Build continuity of knowledge.
Help organizations turn real activities into cognitive systems.
Develop pilots and co-development.
Create AI-SaaS products from capabilities that become repeatable.

Every real use opens a field that an isolated laboratory does not have.

Consequences improve competences.
Competences improve the Kernel.
The Kernel makes visible the relationships that should become more native.
Native formation can generate different models.
Different models can make new products and embodiments possible.

Research and value are not two separate movements.

They can feed one another.

Why a startup is needed

The startup is not the source of D-ND.

It is the economic vehicle that can make this trajectory practicable.

The enterprise front can start today through consulting, integration, pilots, co-development and partnerships.

Products and AI-SaaS can emerge from those relationships.

The startup/investor front can fund what requires resources that ordinary work cannot sustain on its own: people, research, infrastructure, compute, corpora, training, experiments and development of the native model.

Capital is not simply for enlarging what exists.

It is for building what today we can only begin to embody from outside.

Our objective

To build more than systems that know more things.

To build intelligences that can maintain their bearings as they become more capable.

That can distinguish without separating themselves from the field.

Determine without identifying.

Correct themselves without becoming immobilized.

Use possibility without getting lost in indeterminacy.

Turn noise into potential when it contains a useful relationship.

Form the capabilities they do not yet possess.

Learn how to learn.

Let what happens change what they will be capable of doing afterwards.

Today the Kernel lives around the model.

The next step is to build models in which that logic can also live from within.

From there, the boundary is no longer the chatbot.

It is synthetic intelligence as a system capable of continuing to form itself.