AI Cumulative Hiatus Delta: The Missing Test for Cumulative Intelligence

The Missing Test for Cumulative Intelligence

The modern (ai) industry has become extraordinarily good at demonstrating capability.

Models can write, reason, code, analyse, plan, retrieve, classify, summarise, operate software and increasingly act through tools.

Every generation appears more capable than the last.

Benchmarks rise.

Context windows grow.

Agents become more autonomous.

World models become richer.

Formal verification is increasingly attached to generative systems.

And from this progression, one very large conclusion is repeatedly drawn:

human-level artificial intelligence is approaching.

Perhaps within years.

Perhaps sooner.

But there is a problem with this conclusion.

Increasingly intelligent-looking output is not the same thing as cumulative intelligence.

And if the ambition is genuine human-replacing intelligence, that distinction is fundamental.

The missing capability can be described as the:

AI Cumulative Hiatus Delta

AI CHD.

It is the missing capability between:

probabilistic generation and dependable cumulative knowledge continuity.

In simpler terms:

Current (ai) can create the next version. It has not yet demonstrated how it independently preserves what must remain true while changing only what legitimately needs to change.

That may sound like a subtle difference.

It is not.

It may be the difference between an extraordinarily powerful tool and a genuinely cumulative intelligence.

Succession is not continuity

Suppose a system produces a sequence of states:

S_1, S_2, S_3, S_4, …

Each new state may be highly plausible.

It may even be better than the previous one.

But plausibility alone does not establish that:

S_1 → S_2 → S_3 → S_4

represents the legitimate evolution of one coherent state.

Something may have been lost.

A dependency may have disappeared.

An obligation may have silently changed.

A previous decision may have become inconsistent with a new one.

An assumption may have been forgotten.

A fact may have been compressed into something statistically convenient but semantically wrong.

The important question is therefore not:

Does the next state look intelligent?

It is:

What makes the next state a lawful continuation of the previous one?

This is where AI CHD begins.

Kant saw the structural problem long before computers

The philosophical roots of this problem reach back to Immanuel Kant.

In the Critique of Pure Reason, first published in 1781 and substantially revised in 1787, Kant was concerned with the conditions that make coherent experience possible.

Human experience is not merely a stream of disconnected impressions.

Objects persist.

Events have relationships.

Past experience constrains present understanding.

Representations belong together within a unified experience.

The world changes, yet we continue to understand it as the same world.

Kant's concern was therefore not simply the production of representations.

It was their unity.

The structural question can be translated into modern engineering language:

What makes changing representations belong to one coherent, continuing state?

Current (ai) is exceptionally capable at generating representations.

But the fact that those representations appear coherent does not by itself establish the conditions under which they are cumulatively continuous.

This is why the Kantian analogy remains useful.

The issue is not consciousness.

It is not whether a machine has a soul.

It is not whether the machine "feels" anything.

The operational question is much harder to evade:

Where is the mechanism of continuity?

The Scaffolding Paradox

The strongest evidence for AI CHD may not be failure in current research.

It may be the direction of the research itself.

As generative systems become more capable, researchers increasingly surround them with additional machinery providing properties that probabilistic generation alone does not reliably supply.

We see:

symbolic systems supplying law and formal constraint;

world models supplying richer representations of physical persistence;

state-space architectures supplying recurrent state;

vector databases and retrieval systems supplying external memory;

formal verification supplying proof and admissibility checking;

agent frameworks supplying planning, tools and persistent workflows.

None of those developments is evidence that (ai) research is failing.

Quite the opposite.

They are important engineering advances.

But they reveal something important.

The properties required for dependable cumulative intelligence are not currently being demonstrated simply as a consequence of scaling probabilistic generation.

They are increasingly being supplied through specialised complementary systems.

That gives us the:

Scaffolding Paradox

The more capable current (ai) becomes, the more dependable operation appears to require additional machinery for state, grounding, memory, verification, admissibility and continuity.

This does not prove that a future architecture can never integrate these capabilities.

It does something more defensible.

It demonstrates that generation and cumulative intelligence are not currently the same property.

AlphaGeometry and the problem of supplied law

Neuro-symbolic systems provide a useful example.

A system such as AlphaGeometry combines generative search with formal symbolic deduction.

The probabilistic component can propose possible paths.

The symbolic system determines whether those paths satisfy the rules of geometry.

This is enormously powerful.

But notice what has happened.

The constitution already exists.

The geometry is formalised.

The admissible transformations are supplied by the mathematical domain.

The system is effectively operating within a predefined family of permissible transformations:

Γ_K

where K represents the governing structure of the domain.

The system can test:

φ ∈ Γ_K ?

Is this transformation allowed?

That works beautifully in domains where the rules can be made explicit.

But open-world intelligence faces a harder problem.

What happens when the rules themselves change?

The Γ_K problem

Suppose an organisation operates under some set of governing constraints:

Γ_K

These may include technical requirements, laws, customer commitments, physical constraints, financial limits, ethical obligations and accumulated organisational decisions.

The system receives a proposed transformation:

φ

and tests whether:

φ ∈ Γ_K

This is already difficult.

But real intelligence encounters something worse.

Reality changes.

A customer changes strategy.

A regulation changes.

A new scientific discovery invalidates an assumption.

A supplier fails.

An unexpected engineering condition appears.

A business objective disappears.

Now the governing structure itself may need to move from:

Γ_K  to  Γ_K'

This is Admissibility Evolution.

And here the real problem appears:

By what mechanism does the system establish that the change from Γ_K to Γ_K′ is itself legitimate?

Generating a new rule set is easy.

Current (ai) can generate alternative policies, strategies and architectures all day.

The challenge is different.

It must establish:

which previous obligations remain binding;

which have been legitimately superseded;

which dependencies must survive;

which assumptions can now be discarded;

which new constraints become authoritative;

and critically:

why this transformation remains a legitimate continuation of the previous state.

That is far more demanding than producing another plausible frame.

The lightbulb problem

This distinction can be illustrated with IT Dave.

Dave is asked:

Change the lightbulb.

A sufficiently capable system may be able to identify the bulb, order a replacement, schedule the work, analyse energy use, calculate cost and eventually instruct a robot to replace it.

It may also generate creative alternatives.

Use LEDs.

Install skylights.

Move the workstation.

Use natural light.

Remove the room.

That does not yet establish the deeper capability.

The interesting Dave question is:

Does it have to be a lightbulb?

The important part is not the sentence itself.

Any modern model can generate it.

The important part is whether Dave has recognised that the original governing frame is no longer valid.

Perhaps the actual requirement was illumination.

Perhaps even illumination is no longer the requirement.

Perhaps the process that created the requirement has changed.

Dave is not merely exploring alternatives inside a statistical possibility space.

He is changing:

Γ_K → Γ_K'

and carrying the legitimate remnants of the old world into the new one.

That is cross-domain continuity.

Statefulness is not continuity

This distinction becomes particularly important when considering state-space models.

Architectures such as Mamba are designed to carry compact recurrent state across long sequences.

That is valuable.

But a compressed hidden state is not the same thing as an authoritative history.

And continuity does not require remembering everything.

Human beings do not remember everything either.

The harder requirement is:

Can the system establish what may safely be forgotten?

Suppose some piece of information x exists in the historical state.

Forgetting it is itself a transformation:

forget(x)

The important question becomes:

forget(x) ∈ Γ_K ?

Is forgetting x admissible?

Forget what somebody had for lunch six years ago and nothing important may happen.

Forget that an apparently obsolete database field exists because it supports a regulatory obligation and the organisation may later fail an audit.

The difference is not simply memory capacity.

It is admissible forgetting.

A cumulative intelligence must preserve everything whose loss would invalidate the legitimate continuation of state.

That is a significantly harder problem than maintaining a longer computational "now."

World models and grounding

World models attack another important deficiency.

Architectures such as JEPA-style systems attempt to learn latent structure rather than merely predict surface-level observations.

This may enable stronger representations of physical persistence, object behaviour and future world states.

That matters.

A system which understands that an object continues to exist behind an obstruction has a richer relationship with physical reality than a model performing simple token succession.

But predictive grounding is not yet authoritative continuity.

A model can predict what the world is likely to do without establishing:

which entity remains the same entity;

which properties legitimately changed;

which relationships must persist;

which previous obligations remain binding;

or

why a new state should supersede the old one.

World modelling therefore attacks part of the grounding problem.

It does not by itself close AI CHD.

Formal verification and the boundary of certainty

Formal verification comes closer to something like lawful continuity.

If a system is supplied with precise axioms, definitions and constraints, a theorem prover can verify whether a proposed transformation follows from them.

That is exceptionally powerful.

But formal verification still depends on one fundamental condition:

The constitution must already have been represented.

The verifier can determine:

φ ∈ Γ_K

but the real world continually asks:

Γ_K → Γ_K'

And even before that, somebody must determine whether the formal representation of K corresponds adequately to reality.

This is where the boundary between formal systems and open-world intelligence remains difficult.

Mathematics can prove a theorem from axioms.

It cannot by itself establish that the axioms represent the correct problem.

The Composition Problem

Suppose we solve all these component problems independently.

We build:

a brilliant generator;

a perfect world model;

a persistent state system;

a formal verifier;

a symbolic reasoning engine;

a powerful planner;

a reliable action system.

Have we created cumulative intelligence?

Not necessarily.

Because intelligence must survive the transformations between them.

Imagine:

Reality
   ↓
Perception
   ↓
World Representation
   ↓
Memory
   ↓
Reasoning
   ↓
Verification
   ↓
Planning
   ↓
Action
   ↓
Changed Reality

Every arrow is a possible hiatus.

Meaning can change.

Ontology can change.

Context can disappear.

Authority can be lost.

Evidence can be summarised incorrectly.

A locally correct subsystem can hand a semantically distorted state to another locally correct subsystem.

Anyone who has built large distributed systems understands this immediately.

Local correctness does not guarantee global correctness.

Therefore:

component continuity ⇏ system continuity

This is the Composition Problem.

Even if each scaffold works perfectly, the composite system must still demonstrate that cumulative continuity survives across the entire architecture.

Otherwise AI CHD has simply moved from inside the model to the interfaces between the components.

The burden of proof should change

This leads to a different way of evaluating claims of AGI.

We do not need to argue that AGI is impossible.

We do not need to claim transformers can never achieve it.

We do not need to invoke consciousness.

We do not need to claim humans possess some mystical property machines can never reproduce.

We can ask something much simpler:

Show us the mechanism.

If a system is claimed to possess cumulative intelligence, it should be possible to test for the relevant properties.

This suggests an:

AI CHD Audit

1. Continuity Discrimination

Given:

S_t → S_t+1

how does the system establish that S_t+1 is a legitimate continuation rather than merely a plausible successor?

2. Admissible Forgetting

How does the system determine whether:

forget(x) ∈ Γ_K

rather than simply discarding information because it appears statistically unimportant?

3. Cross-Domain Continuity

When information moves between domains:

D_1 → D_2 → D_3

how are meaning, authority, dependency and invariant structure preserved?

4. Admissibility Evolution

When reality changes, how does the system establish:

Γ_K → Γ_K'

and demonstrate that changing the governing rules is itself a legitimate continuation?

5. Invariant Accounting

Following a state transition, can the system explicitly account for:

K_preserved
K_superseded
K_new

and provide evidence for why each element belongs in the corresponding category?

6. Novelty Legitimacy

When the system encounters something not adequately represented by its existing governing structure, how does it distinguish:

genuine novelty

from

error, noise, hallucination or adversarial input?

And if novelty is accepted, how does it incorporate that discovery without silently invalidating legitimate prior state?

This may be the hardest test of all.

Why benchmarks are insufficient

Most existing benchmarks measure performance.

Can the system solve the problem?

Can it answer correctly?

Can it complete the task?

Can it code?

Can it reason through a puzzle?

Can it operate software?

Those are valuable measurements.

But they are not tests of cumulative intelligence.

A system can achieve spectacular task performance while repeatedly rebuilding its working state from context.

It can perform brilliantly inside supplied constraints without independently establishing whether those constraints remain legitimate.

It can retrieve history without understanding which parts of history must remain binding.

It can generate alternatives without establishing when an alternative should become authoritative.

Benchmark intelligence therefore risks measuring:

how intelligently the system behaves inside the current frame

rather than:

whether the system can lawfully maintain and evolve the frame itself.

Those are different capabilities.

Save IT Dave

This is why IT Dave matters.

Dave is not special because humans have perfect memory.

They do not.

Dave is not special because humans never make mistakes.

They do.

Dave is not special because humans always reason logically.

They plainly do not.

Dave matters because organisations currently use humans to close the continuity gaps between systems, domains and changing circumstances.

Dave remembers why one constraint exists.

He knows when another can be discarded.

He notices when Finance has changed something that invalidates Engineering's assumption.

He understands when the customer's new strategy makes last year's technically perfect solution irrelevant.

He knows when the written procedure no longer represents reality.

And occasionally he changes the frame itself.

Dave currently acts as a continuity authority.

That does not mean machines can never perform the same function.

It means that anyone claiming to have replaced Dave should demonstrate that capability rather than infer it from increasingly impressive outputs.

The real challenge to AGI claims

The strongest criticism of current AGI narratives is therefore not:

"LLMs are only autocomplete."

They plainly do far more than the phrase usually implies.

Nor is it:

"machines can never think."

That is an unnecessary philosophical claim.

The stronger challenge is:

Increasingly intelligent output does not itself demonstrate cumulative intelligence.

And therefore:

AGI is not demonstrated by producing increasingly intelligent outputs. It is demonstrated when a system can preserve, revise and justify the continuity under which those outputs belong to one cumulatively coherent intelligence.

That moves the discussion away from appearance.

Away from hype.

Away from benchmark theatre.

Away from vague statements about emergence.

And back to architecture.

Show the invariant.

Show the continuity mechanism.

Show admissible forgetting.

Show the lawful frame change.

Show the cross-domain mapping.

Show the composition.

Show why the resulting state remains legitimate.

If the mechanism exists, demonstrate it.

If it does not, then the burden of proof has not been met.

The reclassified future

None of this diminishes the extraordinary importance of current (ai).

The opposite may be true.

We may be building the most powerful technological scaffolding human beings have ever created.

Generation.

World modelling.

Simulation.

Search.

Verification.

Automation.

Formal reasoning.

Robotics.

Massive information retrieval.

Those systems could transform almost every field of human activity.

But enormously capable scaffolding is not automatically cumulative intelligence.

And confusing those two things creates both technical and economic danger.

The danger appears when organisations remove the human continuity authority before the replacement architecture has demonstrated the required property.

Then the organisation may possess a spectacular collection of components:

a probabilistic mouth,

a predictive world representation,

an external memory,

a symbolic constitution,

a formal verifier,

an agent runtime,

and a mountain of stored context.

Yet still need somebody to determine:

what remains true;

what has changed;

what can safely be forgotten;

which obligations survive;

when the constitution itself must change;

and

why.

For the moment, that person is still Dave.

And Dave is still asking:

Does it have to be a lightbulb?

That is not an argument against (ai).

It is a challenge to the claim that the journey from extremely capable generation to cumulative human-replacing intelligence has already been explained.

It has not.

The AI Cumulative Hiatus Delta remains open.

And until the industry can demonstrate how it crosses it, perhaps we should stop treating Dave's extinction as a settled engineering outcome.

Save IT Dave.

Not because machines are weak.

Because extraordinary capability is not yet the same thing as dependable cumulative intelligence.