A Derivation of Intelligence from Guided Relation-Creation

Intelligence is one of those words that gets harder to pin down the more carefully you look at it. Different traditions ground it in different places — computation, language, reasoning, planning, behavior — and the answers don’t converge. Whatever intelligence is, it doesn’t have an obvious natural boundary, and trying to define it directly mostly tells you about whoever’s doing the defining.

So this won’t start with intelligence. It starts somewhere simpler and works up.

The argument exists in two forms. There’s a full mathematical derivation that works through everything with lemmas and proofs. This is the same argument in prose — same moves, less notation, organized around the questions that actually drive it forward.

Why start with viability rather than intelligence?

Because intelligence isn’t the first problem. Life is.

Reality exists. Things exist inside it. But existing doesn’t guarantee continuing to exist. Finite organized systems can fail — they can be broken, starved, poisoned, dissolved, ruptured, disordered, destroyed. A finite organized system continues only when certain conditions hold, and those conditions aren’t optional. They’re what it means for the system to remain itself.

This is more basic than intelligence, and the framework can build upward from it. Whatever intelligence turns out to be, it has to emerge somewhere on the path from finite-organization-that-keeps-going to whatever the upper levels are. We just have to walk it.

What does it mean for a system to continue?

A cell is the clearest starting case. It exists inside reality, fully subject to physics, not exempt from anything. It continues only while certain conditions hold: membrane integrity, usable energy, material exchange, internal chemical order, repair, protection from things that would destroy it.

This is more specific than “the cell exists.” Life is maintained existence — and the maintenance is the cell’s own work. A rock also exists. A rock persists. But a rock’s persistence isn’t performed by the rock; it happens to the rock, as long as the physics happens to hold. The cell is different. The cell does what’s required for its continuation. Its persistence is its own activity.

So the first distinction we need is between persistence that happens to a system and persistence the system performs. Finite organized systems that perform their own persistence are the subjects of what follows.

Call the conditions whose continuation makes the system the kind of thing it is its organization. For a cell, that includes the things named above. For other kinds of systems, it includes their own analogues. We stay general about what specifically counts as a system’s organization, because that varies with the kind of system — but we commit to there being some such structure for any organized system the argument is about.

The system’s viability depends on its relations with reality. Some relations preserve its organization. Others damage it. This distinction is more basic than moral right and wrong, more basic than rational true and false. It’s the primitive difference between what sustains a system and what destroys it. Everything that follows builds on this distinction.

When does motion become action?

A rock can be moved by the world. Wind wears it down. Water carries it. Gravity pulls it from a hill. The rock enters new relations. But the rock doesn’t produce those relations. It’s moved into them.

A living system can produce new relations from itself. A cell opens or closes channels. It absorbs or expels. It repairs. It moves toward or away from things. It divides. Its own organization participates in producing the relation-change.

This is what action is, in the framework’s sense: relation-change generated by the system’s own organization. A stone rolling downhill moves but doesn’t act, because the moving isn’t done by the stone. A cell moving toward nutrient acts, because the moving is what the cell does.

The distinction matters because once a system can act, its activity becomes part of the conditions of its own continuation. Before action, the system is exposed to what reality does to it. After action, it’s also exposed to what it does into reality. It can bring itself into sustaining relations or destructive ones. Toward nutrient or toward poison. Toward repair or toward worse damage.

Why must action be guided?

If the system produces viability-relevant relations through action, and if the same action can sustain in one situation and destroy in another, then action can’t be arbitrary with respect to reality. If nutrient and toxin make no difference to what the system does, it can’t reliably approach one and avoid the other. If tolerable warmth and lethal heat make no difference to what it does, action can’t reliably preserve it.

So action has to be guided. Guidance means viability-relevant differences in reality make a difference to what the system does. The system’s action varies with what matters for its continuation.

This is the first major move. A living system isn’t merely organized. It’s organized in such a way that its own relation-changing activity has to remain sensitive to the reality that sustains or destroys it.

Life is guided relation-creation.

This is a definitional commitment, not a logical consequence. We’re characterizing life by the structural property the argument just showed is necessary for reliable viability. Choosing this definition is what lets everything that follows build.

What changes when guidance becomes mediated by a model?

Guidance can work two ways.

The simple way is immediate. A chemical gradient changes the cell’s internal state; the altered state changes movement. A damaged membrane triggers repair. A concentration difference opens or closes exchange. Reality couples directly to activity, and activity responds. The system doesn’t need to represent the world, think about it, or carry anything across time. It only needs viability-relevant reality to make a difference to what it does.

Immediate guidance is enough for some forms of life. But it’s bounded by what’s happening right now. It can’t explain how a system acts on something absent, remembered, anticipated, or not currently in contact.

For that, guidance has to become mediated. The system carries traces of past contact. It can be altered by past nourishment, past danger, past injury, past success, past failure. It can compare present to retained. Prepare for what hasn’t happened yet. Act on what it’s learned rather than only on what’s pressing on it right now.

When that happens, action isn’t guided just by immediate coupling. It’s guided through a carried structure that lets the system act beyond present contact.

Call that carried structure a model.

A model in this sense is broad. Not necessarily a picture in the head. Not necessarily explicit. Not necessarily symbolic. Not necessarily conscious. A model is any retained, action-guiding structure that lets a system act from more than what’s immediately present. It might be a learned sensitivity, a memory, an expectation, a spatial map, a body schema, a category, a plan, a prediction, a theory, an institutional record.

The threshold for counting as a model: the carried structure has to preserve, organize, anticipate, or generalize viability-relevant reality beyond immediate coupling. A persistent state that just sits there isn’t a model. A structure that lets the system act on the world differently because of what it carries is.

Why does a model create a new kind of danger?

A model is powerful because it frees action from the immediate. The system can act on what’s absent. It can remember. Prepare. Anticipate. Avoid danger before impact. Seek what isn’t currently touching it.

But the same power opens a new structural problem. The model is not reality. Action is selected from the model — from the carried guidance-state — but the action’s actual consequence happens in reality. So the source of action and the field of consequence aren’t identical. The system acts from what it carries. What it does happens in what is real.

When what the system carries loses contact with what’s real, action can stay coherent inside the model while becoming destructive in reality. A stale model can guide action. A narrow model can guide action. A false model can guide action. An overconfident model can guide action. A model protected from correction can guide action. The system acts intelligibly from its representation while its action produces relations the model didn’t anticipate or refuses to incorporate.

This danger is specific to systems that act through models. Immediate guidance doesn’t have it — there’s no gap to grow, because the system is coupled to reality in real time. Mediated guidance has it built in, because the model persists and can be acted from even when reality has changed underneath it.

So model-guided action requires more than a model. It requires contact between model and reality.

What does it mean for the model to remain in contact with reality?

Contact isn’t the model containing all of reality. No finite model can. Contact isn’t certainty. A system can be uncertain and still be in contact. Contact isn’t the model equaling the world — that wouldn’t even be coherent for a finite system inside a larger reality.

Contact is this: viability-relevant differences in reality can still constrain the system’s guidance. When the world differs in ways that matter, those differences can still reach the system’s guidance, the action that follows from it, or the system’s confidence in that action. When that fails, the model is sealed off, and action continues from a representation the world has already falsified.

Put directly: for any difference in reality that matters for action, the system’s guidance has to be able to register a corresponding difference somewhere — in what it represents, in what it does, or in how much it trusts what it does. If it can’t register that difference, the model is acting blindly with respect to that aspect of reality, even when everything seems fine from the outside.

That’s the third major commitment of the argument. Model-guided action remains reality-responsive only while the model stays in contact with reality. The next question is where contact can break.

Where can contact fail?

To find where, we need the minimal structure of model-guided action. Stripped down, it’s a cycle. Reality, the system’s guidance-state, the action guidance selects, the relation that action produces in reality, the system’s later guidance-state — and running alongside the cycle, the system’s estimate of how far its current guidance can be trusted.

Each step in this cycle is a place where viability-relevant reality has to be preserved. Each is a place where preservation can fail.

The first failure-site is entry. Relevant reality has to be able to get into the guidance-state. If a viability-relevant difference exists in the world but doesn’t produce a difference in the model, the system can’t act on it. A toxin is present but isn’t registered. A danger appears but no sign of it enters. A contradiction exists but no evidence of it reaches the model. If entry fails, action starts from a world the system hasn’t received.

The second failure-site is field. Entry isn’t enough. A relevant difference might enter the system and still be placed outside the field the model treats as relevant — counted as background, noise, exception, someone else’s problem. The model represents food and ignores the predator. It represents immediate reward and ignores later exhaustion. It represents production and ignores waste. The system has organized relevance too narrowly. If field fails, the system acts into more reality than it has modeled.

The third failure-site is return. Action changes reality. The relation it produces has to come back to the system as usable guidance. The cell moves but doesn’t register the difference between nourishment and damage. The institution acts but only tracks output, not harm. The model predicts but the result never makes it back. If return fails, action keeps happening while losing contact with what it actually does.

The fourth failure-site is revision. Consequences can return without changing the model. Damage comes back, behavior doesn’t change. Contradiction comes back, the system files it as noise. Correction comes back, gets stored, doesn’t actually update the structure that guides action. Return asks whether the result comes back. Revision asks whether what comes back can change future guidance. If revision fails, the system is informed without being corrected.

The fifth failure-site is measure. Even a model whose entry, field, return, and revision are working can mismeasure its own incompleteness. Weak evidence, high confidence. Strong evidence, treated as unusable. Knowing something locally and applying it globally. The problem isn’t just that the model is incomplete — it’s that the system mismeasures its incompleteness. If measure fails, action is no longer proportioned to the contact the system has actually earned.

So: five sites. Entry, field, return, revision, measure.

Why does this structure close at five and not more?

The five weren’t chosen by preference. They came out of the cycle.

Reality has to affect guidance — that’s entry. Guidance has to include the field its action enters — that’s field. The relation produced by action has to be available to future guidance — that’s return. Returned relation has to be able to alter the guidance-state — that’s revision. The system has to estimate the strength and limit of its guidance — that’s measure.

So a proposed sixth contact-site has only two possible forms. Either the minimal cycle is missing a necessary term, or one of the existing terms contains another irreducible contact-relation hiding inside it. If neither is true, the proposal isn’t a sixth contact-site.

Lots of things matter for model-guided systems besides these five.

Memory matters for any guidance to persist across time — but memory isn’t a contact-relation. It’s what the existing contact-relations require to operate beyond an instant.

Execution fidelity matters for whether the chosen action actually gets performed — but execution is implementation, not contact.

Goal-formation matters for what action gets chosen — but goals direct rather than constrain reality-contact.

Attention matters for what gets brought into the model — but attention is a selection mechanism within entry and field.

Power matters because it scales effect — but power multiplies contact or its absence rather than constituting it.

Each of these belongs to a neighboring structure. Memory to persistence. Execution to implementation. Goals to direction. Attention to selection. Power to capability. None of them adds a new place where viability-relevant reality has to be preserved if action is to stay reality-guided. They surround the contact-structure. They don’t expand it.

That’s what closure means here. The five sites aren’t a complete account of everything that matters for an intelligent system. They’re the complete account of where model-world contact can fail. The structure is closed against expansion of that specific question. Other questions are real but need their own structures.

What does intelligence become, given all of this?

Walking back through the path: finite organization that performs its own maintenance; action as relation-change the system produces from itself; guidance as the requirement that action varies with reality; immediate guidance for some forms of life; mediated guidance through carried structures called models; the model-gap as the new danger that mediation creates; contact as what closes that gap; and the five sites where contact has to be maintained.

Intelligence is now sayable in the framework’s terms:

Intelligence is contact-closed model-guided relation-creation.

Plainer: intelligence is the capacity of a finite system to guide its relation-changing action through a model while keeping that model answerable to the reality where the action lands.

This doesn’t require carbon. Or neurons. Or language. Or consciousness as a starting assumption. It requires the structural features — model-mediated guidance and the maintenance of contact across the five sites. So it applies to a cell whose chemistry implements all this, to a person whose nervous system implements it, to an institution whose procedures implement it, to an artificial system whose architecture implements it. Wherever the structure obtains.

What this definition makes visible is what intelligence isn’t.

Intelligence isn’t the production of effects.

A system can be enormously capable — manipulating symbols, optimizing toward targets, generating output, acting quickly and forcefully — and still fail to be intelligent in this sense. If reality can’t enter its guidance, or if it can’t include the field its action affects, or if consequences don’t return, or if returned reality can’t revise it, or if it can’t measure the limits of its own contact, then it’s acting in the world without keeping its action answerable to what the world actually is. That’s capability without closed contact.

The distinction is sharp once it’s stated. Capability is the power to produce effects. Intelligence is the power to produce effects while remaining in contact with what those effects do. A capable system without contact produces effects rapidly and confidently — blindly with respect to the field its action enters. A capable system with contact produces effects whose consequences inform what it produces next. The first is power without intelligence. The second is power that intelligence is keeping faithful.

This matters in practice. Many of the systems with the most effect in the world right now — markets, platforms, large institutions, scaled algorithms — are extremely capable. Whether they’re intelligent in this sense is a separate question. The question is whether their actions stay in contact with what their actions actually do, across all five sites. Whether the relevant reality can enter their guidance. Whether their representation of what they affect includes the actually affected field. Whether consequences return. Whether returned consequence can change their structure. Whether they accurately measure how much of their contact they’ve actually earned.

The framework doesn’t answer that for any specific system. But it tells you what to look for. And where contact most often breaks. And why a system losing contact at any of these sites becomes more powerful and more blind at the same time.

That’s what intelligence is. And that’s what its absence looks like when capability continues without it.

---

1. Starting Point

The aim is to understand intelligence from the ground up.

Begin with reality and finite organization.

Reality exists. Finite organized systems exist within it. But the continued existence of a finite organized system is not guaranteed by the mere fact that it exists. A system can continue, or it can fail. It can remain organized, or it can be broken, starved, poisoned, dissolved, ruptured, disordered, or destroyed.

So the first problem is not intelligence.

The first problem is viability.

$$\boxed{ \text{How can a finite organized system continue within a reality that does not guarantee its continuation?} }$$

A living cell is the clearest starting case. It exists inside reality. It is not outside physics. It is not protected from the world by definition. It remains itself only while certain real conditions hold: membrane integrity, usable energy, material exchange, internal chemical order, repair, and protection from destructive conditions.

So life is not mere existence.

Life is maintained existence.

A living system persists only while its organization remains viable within reality.

---

2. Finite Organization

Let:

$S$

be a finite organized system.

Let:

$W$

be the relevant world-condition in which $S$ exists.

The system is finite because:

$S \neq W$

It is not identical with the whole of reality. It has limited access, limited structure, limited power, and limited persistence.

The system is organized because there is some structure:

$O(S)$

whose continuation distinguishes the system remaining itself from the system failing, dissolving, or becoming something else.

For a cell, $O(S)$ includes boundary, metabolism, repair, internal regulation, and exchange with the environment.

---

3. Viability

Define viability as the continued preservation of the system’s organizing conditions.

Let:

$V(S,W)=1$

mean that $S$ remains viable in world-condition $W$.

Let:

$V(S,W)=0$

mean that $W$ damages, destabilizes, or ends the organization of $S$.

So viability is not an optional value added from outside. It is the condition under which the system continues as the system it is.

$$\boxed{ \text{Viability = continued preservation of organizing conditions.} }$$

A finite organized system may be viable under one world-condition and non-viable under another.

So there can be:

$V(S,W_1)=1$

and:

$V(S,W_2)=0$

This means reality makes a difference to whether the system continues.

---

4. Relations

A finite system does not exist apart from reality. It exists in relation to reality.

Let:

$Rel(S,W)$

denote the relation between system $S$ and world-condition $W$.

Some relations preserve or restore the system’s organization.

Other relations damage or destroy it.

So we can define:

$Rel(S,W) \text{ is viability-supporting}$

if it contributes to:

$V(S,W)=1$

and:

$Rel(S,W) \text{ is viability-damaging}$

if it tends toward:

$V(S,W)=0$

This gives the first primitive distinction:

$\text{good-for-continuation}$

and:

$\text{bad-for-continuation}$

This is not moral good and bad. It is not yet rational good and bad. It is not yet true and false.

It is the deeper distinction beneath all of them:

$$\boxed{ \text{Some relations sustain the system; others destroy it.} }$$

---

5. First Lemma: Viability Requires Relation-Fit

Lemma

If a finite organized system can continue in some world-conditions and fail in others, then its continuation depends on its relation to reality.

Proof

Let $S$ be a finite organized system.

Assume there exist two world-conditions:

$W_1$

and:

$W_2$

such that:

$V(S,W_1)=1$

and:

$V(S,W_2)=0$

So $S$ continues under $W_1$, but fails under $W_2$.

Since $S$ exists in relation to each world-condition, the relation:

$Rel(S,W_1)$

supports continuation, while:

$Rel(S,W_2)$

does not.

Therefore, the system’s continuation depends on relation-fit with reality.

$$\boxed{ \text{A finite organized system continues only while its relation to reality preserves its organization.} }$$

---

6. Action

A rock also exists in reality. A rock also enters new relations. Wind can wear it down. Water can carry it. Gravity can pull it from a hill. Ice can crack it. Impact can break it.

But the rock is moved into new relations.

The living system is different.

A living system remains wholly inside physics, but its own organization can participate in producing relation-change. A cell can open or close channels. It can absorb or expel. It can repair. It can move toward or away. It can divide. It can change its boundary-relation to the environment.

The rock can be brought into new relations by reality.

The living system can bring itself into new relations with reality.

This is where action begins.

Define action as:

$$\boxed{ \text{Action = relation-change generated by the system’s own organization.} }$$

Action is not mere motion.

A stone rolling downhill moves, but it does not act, because the relation-change is not generated by the stone’s own self-maintaining organization.

A cell moving toward nutrient or closing against toxin acts, because its own organization participates in producing the relation-change.

Let:

$a$

be an action available to $S$.

Let:

$R(S,W,a)$

be the new relation between $S$ and $W$ produced by action $a$.

---

7. The Guidance Problem

Once a system can act, its own activity becomes part of the conditions of its continuation.

Before action, the system is exposed to what reality does to it.

After action, the system is also exposed to what it does into reality.

It can bring itself into sustaining relations.

It can bring itself into destructive relations.

It can move toward nutrient or poison. It can open to what restores it or open to what kills it. It can repair a boundary or worsen the breach.

So action creates the guidance problem:

$$\boxed{ \text{How can self-produced relation-change remain compatible with continued existence?} }$$

For action to remain viable, it cannot be arbitrary with respect to reality. It must vary with differences in reality that matter for continuation.

If nutrient and toxin make no difference to what the system does, then the system cannot reliably approach one and avoid the other.

If tolerable warmth and lethal heat make no difference to what the system does, then its action cannot reliably preserve it.

So guidance means:

$$\boxed{ \text{viability-relevant differences in reality make a difference to action.} }$$

Or, in compressed form:

$\Delta W \rightarrow \Delta a$

where:

$\Delta W$

is a viability-relevant difference in reality, and:

$\Delta a$

is an action-relevant difference in what the system does.

---

8. Second Lemma: Action Requires Guidance

Lemma

If a system creates viability-relevant relations through action, and if the same action can be viability-supporting in one world-condition but viability-damaging in another, then the system’s action must be guided by viability-relevant differences in reality.

Proof

Let $S$ be a finite organized system.

Let $a$ be an action available to $S$.

Let:

$R(S,W,a)$

be the relation produced when $S$ performs $a$ in world-condition $W$.

Let:

$Viable(R(S,W,a))$

mean that the produced relation supports the continued organization of $S$.

Assume there are two world-conditions:

$W_1$

and:

$W_2$

such that action $a_1$ is viable in $W_1$, but not viable in $W_2$:

$Viable(R(S,W_1,a_1))$

and:

$\neg Viable(R(S,W_2,a_1))$

So the same action has different viability-status depending on the world-condition.

Now suppose the system cannot differentiate $W_1$ from $W_2$ in any way that affects action.

Then the difference between $W_1$ and $W_2$ cannot guide what the system does.

So the system may produce the same action in both conditions.

But by assumption, that action is viability-supporting in one condition and viability-damaging in the other.

Therefore, the system cannot reliably preserve viability across those conditions unless the difference between $W_1$ and $W_2$ can guide action.

So:

$$\boxed{ \text{Reliable viability requires action to be guided by viability-relevant differences in reality.} }$$

This gives the first major structure:

$$\boxed{ \text{life is guided relation-creation.} }$$

This is a definitional commitment, not a logical consequence. The framework characterizes life by the structural property the lemma has shown to be necessary for reliable viability.

A living system is not merely organized. It is organized in such a way that its own relation-changing activity must remain sensitive to the reality that sustains or destroys it.

The next step is to distinguish immediate guidance from model-mediated guidance.

9. Immediate Guidance

Guidance can be immediate.

A difference in the world may directly alter the system’s activity. A chemical gradient changes the cell’s internal state, and that altered state changes movement. A damaged membrane triggers repair. A concentration difference opens or closes exchange.

In this case:

$\Delta W \rightarrow \Delta S \rightarrow \Delta a$

A difference in reality produces a difference in the system, and that difference produces a difference in action.

This is already reality-guided life.

The system does not need to represent the world as an object. It does not need to think, judge, believe, or predict. It only needs viability-relevant reality to make a difference to what it does.

Immediate guidance is therefore enough for some forms of life. A system can remain viable because its activity is coupled directly to the conditions that sustain or damage it.

But immediate guidance has a limit.

It binds action to what is presently affecting the system. It does not yet explain how a system can act from what is absent, remembered, expected, possible, hidden, delayed, or not yet directly encountered.

For that, guidance must become mediated.

---

10. Mediated Guidance

A system may carry traces of reality across time.

It may retain the result of prior contact. It may be altered by past nourishment, past danger, past injury, past success, or past failure. It may compare present conditions to retained traces. It may prepare for what is not yet present. It may act according to what has been learned, not only according to what is immediately pressing on it.

At that point, action is no longer guided only by immediate coupling.

It is guided through a carried structure.

Let:

$G_t$

be the guidance-state of system $S$ at time $t$.

Then mediated guidance has the form:

$G_t \rightarrow a_t$

or, when present reality also contributes:

$(W_t,G_t) \rightarrow a_t$

The system acts through something it carries.

This carried guidance-state may be simple or complex. It may be a learned sensitivity, a memory, an expectation, a spatial map, a body schema, a category, a plan, a prediction, a theory, or an institutional record.

The important feature is not that the guidance-state is conscious or linguistic.

The important feature is that action is selected through a structure that can preserve something of reality beyond immediate contact.

---

11. Model

We can now define model:

$$\boxed{ \text{A model is a carried guidance-structure through which action is selected beyond immediate coupling alone.} }$$

A model does not have to be a picture in the head.

It does not have to be explicit.

It does not have to be symbolic.

It does not have to be conscious.

A model is any retained, organized, action-guiding structure that lets the system act from more than the immediately present stimulus.

The threshold is important.

Not every physical alteration is a model. A system counts as model-guided when the carried structure mediates action by preserving, organizing, anticipating, or generalizing viability-relevant reality beyond immediate coupling.

So the transition is:

$\text{immediate guidance}$

to:

$\text{model-mediated guidance}$

or:

$\Delta W \rightarrow \Delta S \rightarrow \Delta a$

to:

$G_t \rightarrow a_t$

where $G_t$ carries action-relevant structure across time, absence, possibility, or abstraction.

This is the beginning of the problem of intelligence.

---

12. The Model-Gap

A model is powerful because it frees action from the immediate.

The system can act on what is absent.

It can remember what has happened.

It can prepare for what has not yet happened.

It can avoid danger before impact.

It can seek what is not currently touching it.

It can compare, anticipate, plan, generalize, and correct.

But the same power creates a new danger.

The model is not reality.

$G_t \neq W_t$

The action is selected from the guidance-state:

$a_t = A(G_t)$

but the action produces a relation in reality:

$R_t = R(S,W_t,a_t)$

So model-mediated action contains a structural gap:

$\text{selection occurs from }G_t$

while:

$\text{consequence occurs in }W_t$

The system acts from what it carries, but what it does happens in what is real.

If what the system carries loses answerability to what is real, then action can remain coherent inside the model while becoming destructive in reality.

A stale model can guide action.

A narrow model can guide action.

A false model can guide action.

An overconfident model can guide action.

A model protected from correction can guide action.

So model-guided action requires more than a model.

It requires contact between model and reality.

---

13. Contact

Define contact as the preservation of viability-relevant reality-difference within model-guided action.

$$\boxed{ \text{Contact = preservation of viability-relevant difference between reality and guidance.} }$$

Contact does not mean that the model contains all of reality.

No finite model can.

Contact does not mean certainty.

A system can remain uncertain and still be in contact.

Contact does not mean exact equality between guidance and world.

$G_t = W_t$

is impossible for a finite system.

Contact means that differences in reality that matter for action, viability, and correction can still make a difference to the guidance-state, to the action selected from it, or to the system’s confidence in that action.

In compressed form:

$\Delta W \Rightarrow \Delta G$

where $\Delta W$ is a viability-relevant difference in reality and $\Delta G$ is a relevant difference in guidance.

More generally:

$\Delta W \Rightarrow \Delta(G,a,K)$

where $K$ is the system’s estimate of how far its guidance can be trusted.

So:

$$\boxed{ \text{A model remains in contact with reality when relevant reality can still constrain guidance.} }$$

---

14. Third Lemma: Model-Guided Action Requires Contact

Lemma

If action is selected through a carried guidance-state, and if different world-conditions require different actions for viability, then the guidance-state must preserve the relevant differences between those world-conditions.

Proof

Let:

$W_1$

and:

$W_2$

be two world-conditions.

Let:

$G(W)$

be the guidance-state formed in relation to world-condition $W$.

Let:

$A(G(W))$

be the action selected from that guidance-state.

Let:

$R(S,W,A(G(W)))$

be the relation produced when system $S$ acts in world-condition $W$ from guidance-state $G(W)$.

Let:

$Viable(R(S,W,A(G(W))))$

mean that the produced relation supports the continued organization of $S$.

Assume there is an action $a_1$ such that:

$Viable(R(S,W_1,a_1))$

and:

$\neg Viable(R(S,W_2,a_1))$

So action $a_1$ is viable in $W_1$, but not viable in $W_2$.

Now suppose the guidance-state does not preserve the relevant difference between $W_1$ and $W_2$:

$G(W_1)=G(W_2)$

Since action is selected from the guidance-state:

$A(G(W_1))=A(G(W_2))$

Therefore the system selects the same action in both cases.

But by assumption, the same action is not viable in both cases.

Therefore the system cannot reliably preserve viability across $W_1$ and $W_2$ unless the guidance-state preserves the viability-relevant difference between them.

So:

$$\boxed{ \text{Model-guided viability requires the guidance-state to preserve viability-relevant differences in reality.} }$$

This establishes the need for model-world contact.

A model-guided system does not need perfect representation. It does not need total knowledge. But it must preserve the differences that matter for the action it guides.

Where the model collapses differences that reality makes decisive, the system acts blindly through its own model.

---

15. Where We Are

The path so far is:

$\text{finite organization}$

$\Downarrow$

$\text{viability}$

$\Downarrow$

$\text{relation-fit}$

$\Downarrow$

$\text{action as system-generated relation-change}$

$\Downarrow$

$\text{guidance}$

$\Downarrow$

$\text{model-mediated guidance}$

$\Downarrow$

$\text{contact}$

This gives the next task:

$$\boxed{ \text{identify the structurally distinct places where contact can fail.} }$$

That is where the contact-sites are derived.

16. Contact-Failure

Contact is the preservation of viability-relevant difference between reality and guidance.

So contact fails when such difference is lost, excluded, blocked, distorted, or mismeasured.

Let:

$\Delta W$

mean a viability-relevant difference in reality.

Examples:

$\Delta W = \text{nutrient rather than toxin}$

$\Delta W = \text{safe path rather than dangerous path}$

$\Delta W = \text{action succeeded rather than failed}$

$\Delta W = \text{model confirmed rather than contradicted}$

For model-guided action to remain viable, $\Delta W$ must be preserved well enough to affect guidance.

In the simplest form:

$\Delta W \Rightarrow \Delta G$

But model-guided action is not only a relation between world and model. The model selects action, action produces consequences, consequences must matter for future guidance, and the system must estimate how far its model can be trusted.

So the general form is:

$\Delta W \Rightarrow \Delta(G,a,R,G',K)$

where:

$G = \text{guidance-state}$

$a = \text{action selected from guidance}$

$R = \text{relation produced by action}$

$G' = \text{future guidance-state}$

$K = \text{estimate of contact-strength and limit}$

A contact-failure occurs when a viability-relevant difference in reality fails to constrain the guidance process where it must.

So:

$$\boxed{ \text{Contact-failure = loss of viability-relevant difference in model-guided action.} }$$

The next question is:

$$\boxed{ \text{Where can this loss occur?} }$$

---

17. Minimal Structure of Model-Guided Action

To locate contact-failure, we need the minimal structure of model-guided action.

There is reality:

$W$

There is a guidance-state or model:

$G$

There is an action selected from guidance:

$a = A(G)$

There is the real relation produced by action:

$R = R(S,W,a)$

There is the later guidance-state:

$G'$

And because the guidance-state is finite and incomplete, there is the system’s estimate of how far that guidance can be trusted:

$K(G,a)$

So the minimal structure is:

$W \rightarrow G \rightarrow a \rightarrow R \rightarrow G'$

with:

$K(G,a)$

This is not meant as a temporal sequence only. It is the contact-structure of model-guided action.

Reality must constrain guidance.

Guidance must select action.

Action must produce a real relation.

The produced relation must matter for future guidance.

And the system must estimate the strength and limit of its guidance.

The contact-sites are the places where viability-relevant reality must be preserved in this structure.

---

18. Entry

The first contact-site is the relation from reality to guidance:

$W \rightarrow G$

Relevant reality must be able to enter the guidance-state.

If a viability-relevant difference exists in the world but produces no relevant difference in guidance, then the system cannot act according to that difference.

Formally:

$\Delta W \nRightarrow \Delta G$

A toxin is present, but the system does not register it.

A danger appears, but no usable sign of it enters.

A contradiction exists, but no evidence of it reaches the model.

An injury occurs, but the system receives no signal of damage.

In each case, the world differs in a way that matters, but the guidance-state does not differ with it.

So the first contact-condition is:

$$\boxed{ \text{Relevant reality must be able to enter the guidance-state.} }$$

Call this entry.

At higher levels, entry appears as sensing, perception, observation, evidence, input, or data. But the general structure is simpler:

$\text{what matters in reality must be able to make a difference inside the system.}$

If entry fails, action begins from a world the system has not received.

---

19. Field

Entry is not enough.

A relevant difference may enter the system but be placed outside the field the model treats as relevant.

This is not simple blindness. It is mis-scoping.

The system may receive something, but treat it as background, noise, exception, externality, or outside the action’s concern.

Let:

$D(a,W)$

be the real field affected by action.

Let:

$F_G(a)$

be the field treated as relevant by the guidance-state.

Field failure occurs when:

$F_G(a) \subset D(a,W)$

and the omitted part contains viability-relevant differences.

The model may represent food but omit the predator near it.

It may represent immediate reward but omit later exhaustion.

It may represent production but omit waste.

It may represent the intended target of action but omit the wider field disturbed by action.

In each case, the system does not merely lack input. It has organized relevance too narrowly.

So the second contact-condition is:

$$\boxed{ \text{The guidance-state must include the relevant field affected by action.} }$$

Call this field.

At higher levels, this appears as context, interconnection, dependency, scope, system-boundary, or affected field.

If field fails, the system acts into more reality than it has modeled.

---

20. Return

The third contact-site appears because action changes reality.

The system may begin with adequate entry and adequate field, act, and create a new relation. But the relation actually produced may not return to the system as usable guidance.

The action lands in reality, but its result does not come back.

Let:

$R_t = R(S,W_t,a_t)$

be the relation produced by action.

Return failure occurs when:

$R_t \nRightarrow G_{t+1}$

The system acts, but the consequences of action do not alter future guidance.

The cell moves, but receives no usable difference between nourishment and damage.

The animal acts, but cannot register success, injury, or changed condition.

The institution acts, but tracks only output and not harm.

The model predicts, but the result never reaches the model.

In each case, action has consequences, but those consequences vanish from guidance.

So the third contact-condition is:

$$\boxed{ \text{The consequences of action must be able to return to the system.} }$$

Call this return.

At higher levels, this appears as feedback, consequence-tracking, outcome-monitoring, error signal, experiment, or result.

If return fails, the system can keep acting while losing contact with the relations its own action creates.

---

21. Revision

Return is still not enough.

A consequence may return, but the guidance-state may not be changeable by it.

Reality comes back, but the model does not move.

The system receives damage but does not alter behavior.

It receives failure but protects the old pattern.

It receives contradiction but treats it as noise.

It receives correction but stores it without changing the structure that guides future action.

So return asks:

$\text{Does the result come back?}$

Revision asks:

$\text{Can what comes back change future guidance?}$

Revision failure occurs when:

$R_t \Rightarrow G_t$

but:

$G_{t+1}=G_t$

even though $R_t$ shows that $G_t$ is wrong, incomplete, unsafe, or misweighted.

So the fourth contact-condition is:

$$\boxed{ \text{Returned reality must be able to revise the guidance-state.} }$$

Call this revision.

At higher levels, this appears as learning, updating, correction, plasticity, supersession, repair, or belief-revision.

If revision fails, the system may be informed without being corrected.

---

22. Measure

Even entry, field, return, and revision do not complete the contact problem.

A finite model is never complete. It has limits. It may be strong in one region and weak in another. It may have current contact here, stale contact there, narrow evidence here, broad uncertainty there.

So the system must not only have contact. It must estimate how much contact it has.

Let:

$\sigma(G,W,a)$

be the actual degree of contact or support between the guidance-state and reality relative to action.

Let:

$\hat{\sigma}(G,a)$

be the system’s estimate of that contact.

Measure failure occurs when:

$\hat{\sigma}(G,a) \not\approx \sigma(G,W,a)$

The system may overestimate its contact and overact.

It may underestimate its contact and fail to act.

It may know something locally but apply it globally.

It may have weak evidence but high confidence.

It may have strong evidence but treat it as unusable.

In each case, the problem is not merely that the model is incomplete. The problem is that the system mismeasures its incompleteness.

So the fifth contact-condition is:

$$\boxed{ \text{The system must estimate the strength, reach, and limit of its contact.} }$$

Call this measure.

At higher levels, this appears as confidence, uncertainty, calibration, threshold, source-weighting, or warranted commitment.

If measure fails, action is no longer proportioned to the contact actually earned.

---

23. Contact-Closure

The five contact-sites are:

$\text{entry}$

$\text{field}$

$\text{return}$

$\text{revision}$

$\text{measure}$

Together:

$$\boxed{ Contact(G,W,a)=Entry \land Field \land Return \land Revision \land Measure }$$

This is not a sequence.

Entry is shaped by field.

Field determines what return can count.

Return matters through revision.

Revision changes future entry and field.

Measure governs the confidence and commitment of the whole system.

The structure is a closure, not a line.

The system remains in model-world contact only while all five sites remain operative as one contact-field.

So:

$$\boxed{ \text{Model-guided action remains reality-guided only if relevant reality can enter, be included in the field, return through consequence, revise guidance, and be measured in its strength and limit.} }$$

The next step is to explain why this structure closes: why any proposed further contact-site must either add a missing term to the minimal structure or reduce to one of these sites.

24. Why the Contact-Structure Closes

The five contact-sites were not chosen by preference. They were derived from the minimal structure of model-guided action:

$W \rightarrow G \rightarrow a \rightarrow R \rightarrow G'$

with:

$K(G,a)$

where:

$W = \text{world-condition}$

$G = \text{guidance-state or model}$

$a = \text{action selected from guidance}$

$R = \text{relation produced in reality by action}$

$G' = \text{later guidance-state}$

$K(G,a) = \text{estimate of the strength and limit of guidance}$

A contact-site is a place where viability-relevant reality must be preserved if action is to remain guided by reality.

So a proposed additional contact-site must do one of two things.

It must either show that the minimal structure is missing a necessary term, or it must show that one of the existing terms contains another irreducible contact-relation.

Otherwise, it is not a new contact-site.

It may still be important. It may be necessary for embodiment, execution, ethics, consciousness, social coordination, memory, attention, or power. But it does not add another site of model-world contact unless it identifies a new way viability-relevant reality can fail to guide model-mediated action.

The five sites correspond to the required contact-relations in the minimal structure.

Reality must affect guidance:

$W \rightarrow G$

This is entry.

Guidance must include the real field into which action enters:

$G \leftrightarrow D(a,W)$

This is field.

Action produces a real relation whose result must be available to future guidance:

$R \rightarrow G'$

This is return.

Returned reality must be able to alter the guidance-state when alteration is required:

$G \+ R \rightarrow G'$

This is revision.

The system must estimate the strength and limit of its guidance:

$G,a \rightarrow K$

This is measure.

So the closure claim is:

$$\boxed{ \text{Any failure of model-world contact must occur as failure of entry, field, return, revision, or measure, unless the minimal structure itself is expanded.} }$$

This is not the claim that nothing else matters.

It is the claim that, for model-guided relation-changing action, these are the contact-sites generated by the structure itself.

---

25. The Contact-Closure Theorem

We can now state the theorem.

$$\boxed{ \textbf{Contact-Closure Theorem} }$$

For a finite system whose relation-changing action is guided by a model, action remains guided by reality only if viability-relevant reality is preserved across five contact-sites:

$\text{Entry}$

$\text{Field}$

$\text{Return}$

$\text{Revision}$

$\text{Measure}$

Reality must enter guidance.

The relevant affected field must be included.

The consequences of action must return.

Returned consequence must be able to revise guidance.

The system must measure the strength, reach, and limit of its contact.

Together:

$$\boxed{ Contact(G,W,a)=Entry \land Field \land Return \land Revision \land Measure }$$

A proposed additional contact-site must either identify a missing necessary term in:

$W \rightarrow G \rightarrow a \rightarrow R \rightarrow G'$

with:

$K(G,a)$

or identify an irreducible contact-relation not captured by the five.

Otherwise, it belongs not to contact-closure itself, but to a neighboring structure: execution, embodiment, memory, attention, goal-formation, ethics, consciousness, social organization, or power.

---

26. Intelligence

We can now define intelligence from the structure already built.

The path has been:

$\text{finite organization}$

$\Downarrow$

$\text{viability}$

$\Downarrow$

$\text{relation-fit}$

$\Downarrow$

$\text{action as system-generated relation-change}$

$\Downarrow$

$\text{guided relation-creation}$

$\Downarrow$

$\text{model-mediated guidance}$

$\Downarrow$

$\text{model-world contact}$

$\Downarrow$

$\text{contact-closure}$

This gives:

$$\boxed{ \text{Intelligence is contact-closed model-guided relation-creation.} }$$

Or, in prose:

Intelligence is the capacity of a finite system to guide relation-changing action through a model while keeping that model answerable to the reality in which action occurs.

This definition is substrate-neutral.

It does not require carbon.

It does not require neurons.

It does not require language.

It does not require consciousness as a starting assumption.

It requires model-guided action under contact-closure.

A system may have power without this.

It may produce large effects.

It may manipulate symbols.

It may optimize toward a target.

It may output convincing language.

It may act quickly, fluently, or forcefully.

But if reality cannot enter its guidance, if it cannot include the field its action affects, if consequences do not return, if returned reality cannot revise it, or if it cannot measure the limits of its own contact, then its action is not intelligent in this sense. It is capability without closed contact.

So intelligence is not mere capability.

Capability is the power to produce effects.

Intelligence is the power to produce relation-changing action while remaining answerable to reality.

$$\boxed{ \text{Capability produces effects. Intelligence keeps effect-producing guidance in contact with what is real.} }$$

This completes the contact account of intelligence.