MATHCIV-006 | Mathematics + Civilisation Series | eduKateSG

Quick Read

A student can appear to be advancing while the underlying mathematical system is becoming weaker.

This can happen when harder questions, calculators, formula sheets, tutors, answer keys or artificial intelligence allow the student to produce increasingly sophisticated work while the mathematical operations underneath that work are no longer independently available.

That creates an important distinction:

Advanced performance is not necessarily advanced capability.

The framework used in this MathematicsOS project therefore separates three ideas:

BaseFloor — the mathematical foundations that must remain usable.

P3: Stable Functioning — mathematics the learner can perform accurately, independently, after a delay, across reasonable variations, with checking and recovery.

P4: Protected Frontier Extension — harder problems, unfamiliar applications, modelling, proof, computational tools and AI-supported exploration that extend the learner beyond the stable base.

The rule is simple:

P4 should extend P3. It should not secretly replace it.

This matters increasingly in an age when powerful tools can generate algebra, explanations, graphs, proofs and complete solutions faster than a learner can construct them.

The objective is therefore not to keep students away from advanced tools.

It is to make sure that the frontier grows without hollowing out the foundations underneath it.

The central problem: a learner can move forward and backward at the same time

Mathematics education is usually imagined as a ladder.

Learn arithmetic.

Then algebra.

Then functions.

Then trigonometry.

Then calculus.

Then increasingly difficult applications.

From that picture, advancement appears straightforward: if a student can solve a harder problem today than six months ago, the system must have improved.

But learning is not quite that simple.

A learner can become better at producing an answer while becoming more dependent on the environment that produces it.

The student may:

remember a procedure only while looking at an example;
recognise a familiar worksheet but fail when the presentation changes;
obtain the correct answer with hints but be unable to reconstruct the route alone;
solve quickly immediately after tuition but lose the method two weeks later;
use an AI system to produce sophisticated reasoning that the student cannot subsequently reproduce;
complete advanced tasks while basic algebra remains fragile.

These are not necessarily failures.

Hints, worked examples, calculators and AI can all be useful.

The problem appears when supported performance is mistaken for independent capability.

The MathematicsOS master therefore defines stable P3 functioning as requiring accurate fundamentals, independent working, retention, appropriate route selection, transfer, checking, recovery and execution under realistic conditions. P4 contains frontier activities such as unfamiliar synthesis, modelling, proof, advanced representation, computational exploration and AI-assisted comparison. Crucially, the P4 admission rule says that a tool must not replace the mathematical operation the learner is supposed to be acquiring.

That gives us a more useful model of progress.

1. The BaseFloor

The BaseFloor is the minimum body of capability that the next layer of learning assumes will remain available.

It is not one universal list.

A Primary Mathematics learner, a Secondary Mathematics learner and an Additional Mathematics learner will have different BaseFloors.

Even within Additional Mathematics, the relevant floor changes with the task.

For example, differentiation may depend upon:

algebraic equivalence
→ indices
→ functions
→ symbolic manipulation
→ interpretation of gradient/rate
→ differentiation procedure
→ checking

If a later layer appears to function only because an external tool repeatedly performs one of those earlier operations, the higher performance may conceal a lower-level dependency.

This is why MathematicsOS does not define the BaseFloor as:

everything previously taught.

That would be impractical.

Instead, the relevant question is:

Which earlier capabilities must remain sufficiently stable for the learner to control the current problem?

That is a much smaller and more useful set.

2. Why protecting the BaseFloor does not mean keeping students on easy work

There is an obvious danger in any foundation-first philosophy.

Teachers can keep repairing prerequisites forever.

A student who struggles with algebra can be told:

“You cannot do Additional Mathematics until every part of algebra is perfect.”

But mathematics is not a perfectly linear staircase.

Later mathematics can sometimes deepen earlier mathematics.

Functions can make algebra more meaningful.

Calculus can reveal why the behaviour of a graph matters.

Trigonometry can expose weaknesses in symbolic manipulation that simpler exercises never revealed.

So BaseFloor protection should not become BaseFloor imprisonment.

The better rule is:

repair the smallest active prerequisite set necessary to make productive forward movement possible.

Then advance.

Observe.

Repair again when necessary.

The system becomes:

Sense
→ Repair
→ Advance
→ Test
→ Resense

rather than:

Repair everything
→ eventually allow advancement

This distinction matters for capable students who have a few fragile components but are ready for richer mathematical work.

3. P3: stable mathematical functioning

P3 is the point at which a mathematical capability becomes operationally dependable.

Again, P3 is an eduKateSG/CivilisationOS architectural category, not a standardized level established by a scientific experiment.

The scientific evidence helps specify what stable learning should be tested for. The P3 label is the control architecture built around those findings.

For a capability to be treated as stable, MathematicsOS asks five increasingly demanding questions.

Receipt

Did the learner actually understand what was presented?

Retention

Can the learner reconstruct the idea after time has passed?

Transfer

Can the learner recognise and adapt the idea when the surface form changes?

Independence

Can the learner perform without the tutor, hint chain, worked solution or AI system?

Execution

Can the learner still control the mathematics under realistic time and cognitive demands?

These distinctions matter because learning can fail between any two stages.

A lesson can be delivered without being understood.

Something understood today can disappear by next week.

Something remembered can remain tied to the worksheet on which it was learned.

Something transferable in untimed practice can collapse during an examination.

And something completed with AI may never have become independently available to the learner at all.

4. Why transfer is one of the hardest tests

A particularly important warning comes from research into children performing mathematics in different contexts.

A large 2025 Nature study examined children who regularly performed arithmetic while working in Indian markets and children learning mathematics primarily through school.

Their abilities were strikingly context-dependent.

In one part of the study, working children correctly handled 85% of actual market transactions, while school children performed very poorly in a closely simulated timed market context. Yet school children outperformed the working children on school-like abstract subtraction and division tasks.

The researchers found similarly sharp differences even when they constructed problems intended to bridge familiarity between the groups. They concluded that there was very limited transfer between the applied and academic mathematical environments they studied.

This does not establish that all school mathematics fails to transfer.

It does something more useful.

It demonstrates why we should not automatically infer:

“The learner can do mathematics in Context A, therefore the same capability will appear in Context B.”

For MathematicsOS, this means a stable base must contain bridges, not merely nodes.

A student who can differentiate ten functions when the worksheet says “Differentiation” may possess procedural performance.

A student who can identify when differentiation is needed inside an unfamiliar rate-of-change problem possesses a stronger form of operational control.

That distinction is central to P3.

5. P4: protected frontier extension

Once a sufficiently stable base exists, education should not stop there.

The learner should encounter mathematics that is harder, less predictable and more powerful.

That is P4.

P4 may contain:

unfamiliar problems;
multi-topic synthesis;
modelling;
proof;
computational exploration;
alternative solution routes;
advanced visualisation;
mathematical research questions;
simulations;
programming;
AI-supported comparison;
tool-assisted exploration.

P4 is where mathematics begins to behave less like a worksheet and more like a living intellectual system.

It is also where modern tools can be enormously valuable.

A graphing system can expose structure that would take a long time to sketch manually.

Code can explore hundreds of cases.

Computer algebra can test conjectures.

AI can generate alternative explanations, compare approaches and suggest questions.

The objective should therefore not be to build a wall around human cognition and forbid external tools.

Civilisation advances partly because humans construct tools that extend what one individual can do.

The important question is instead:

What capability should reside inside the learner, and what capability may productively reside in the tool?

That boundary changes according to the learning objective.

6. The substitution test

Suppose a student is learning differentiation.

An AI system may:

explain what the derivative represents;
generate practice questions;
give a hint when the student is stuck;
compare two valid differentiation methods;
check a completed derivative;
perform the entire differentiation automatically.

All six actions may be useful in some setting.

But they are not educationally equivalent.

If the current objective is:

learn to differentiate independently,

then Action 6 substitutes for the capability being trained.

If the objective is instead:

use derivatives inside a sophisticated optimisation model,

automatic differentiation may be perfectly reasonable, provided the underlying capability has already been established to the level required.

This gives us the Substitution Test:

Is the tool extending the learner’s capability, or performing the capability that we are currently trying to build?

If it is performing the target operation, supported output cannot automatically be recorded as independent learning.

7. Why artificial intelligence makes this problem much more important

Generative AI increases the importance of this distinction because it can produce not merely answers but the appearance of reasoning.

A student can now request:

“Explain every step.”

The AI may produce a beautiful explanation.

But whose explanation is it?

The presence of reasoning on the screen does not establish reasoning inside the learner.

A 2025 PNAS field experiment in high-school mathematics illustrates the risk. Students given access to generative-AI assistance could improve their performance while the technology was available, but unrestricted access produced poorer subsequent unaided learning outcomes; the negative effect was substantially mitigated when the AI system was redesigned with tutoring guardrails.

The result should not be translated into:

“AI is bad for mathematics.”

That would be inconsistent with other evidence.

A separate 2025 randomized controlled trial involving 194 eligible undergraduate students in a Harvard introductory physics course compared carefully engineered AI tutoring with established in-class active learning. Students using the AI tutor achieved substantially stronger immediate post-test outcomes in the two lessons studied, while typically spending less time on the learning activity.

The researchers emphasised that their tutor was not simply an unrestricted chatbot. It incorporated carefully prepared instructional structure, expert-written material, scaffolding and detailed solutions; they also explicitly cautioned against assuming that such an approach would outperform classroom learning in every context.

Put the two studies together and an important principle emerges.

It is not scientifically justified to ask only:

Does AI help?

A better set of questions is:

Which AI?
Designed how?
For which learner?
For which mathematical operation?
During which phase of learning?
Measured while the AI is present or after it disappears?
And what remains independently available afterwards?

That is exactly the problem the P3/P4 distinction is intended to control.

8. AI-assisted performance versus AI-extended capability

Consider two students.

Student A

The student cannot form a quadratic equation independently.

They give every question to an AI system, study the resulting solution and copy the method.

Their homework becomes excellent.

Student B

The student can already construct and solve the relevant quadratic equations.

They use AI to:

generate unusual modelling situations;
compare algebraic and graphical methods;
inspect alternative reasoning;
test edge cases;
challenge their own explanation.

Both students are “using AI for mathematics.”

But the architecture is completely different.

For Student A:

AI is occupying part of the BaseFloor.

For Student B:

AI is extending the frontier above the BaseFloor.

The technology may be identical.

The educational effect can differ because the receiver state differs.

This is why blanket rules such as “use AI” or “ban AI” are much less useful than capability-specific rules.

9. The AI-off test

MathematicsOS therefore uses a deliberately simple safeguard:

Turn the tool off.

Not forever.

Long enough to find out what remains.

The master specifies an AI-off capability protocol with five tests. After AI-assisted learning, the learner should be able to explain the governing idea, solve a fresh structurally similar problem without AI, handle a changed representation or surface form, verify the result independently, and repeat at least one qualifying probe after a delay.

The point is not punishment.

It is measurement.

If the learner succeeds, the AI may have helped build capability.

If the learner fails, the correct conclusion is not:

“The student learned it badly.”

The scientifically safer statement is the one embedded in the master:

Assisted performance observed; independent capability not yet established.

That distinction protects both the student and the accuracy of the learning system.

10. The tool-off principle applies far beyond AI

The same test should occasionally be applied to other supports.

Calculator-off

Can the learner still estimate magnitude, signs and plausibility?

Formula-sheet-off

Does the learner understand the relation well enough to reconstruct or recognise it?

Worked-example-off

Can the learner initiate the route independently?

Tutor-off

Can the student decide what to do without a sequence of prompts?

Notes-off

Can the central structure be reconstructed after a delay?

Familiar-format-off

Can the learner identify the same mathematics when wording or representation changes?

These are not arguments for permanently removing support.

They are diagnostic probes.

Support should help construct capability.

It should not make capability impossible to observe.

11. A stronger model of advancement

We can therefore replace the simple model:

Easy → Harder → Hardest

with something much more useful:

Acquire
→ Stabilise
→ Remove Support
→ Retain
→ Transfer
→ Execute
→ Extend

Then:

Extend
→ return and verify the base
→ extend again

This produces a ratchet.

The learner moves outward without allowing the centre to disappear.

We can describe the principle informally as:

Frontier expansion should be accompanied by periodic base verification.

This is architecture, not a universal empirical equation.

Its purpose is to stop increasingly powerful external systems from creating an illusion of growth.

12. P3 does not mean memorising everything

There is another possible misunderstanding.

If P3 means independent capability, does every formula and technique have to be memorised permanently?

No.

Civilisation itself functions through external memory.

Books, tables, software, diagrams, databases and calculators allow human beings to operate beyond biological memory.

Mathematics has always co-evolved with external representations.

The P3 question is therefore not:

“Can the student do everything without any external object?”

The question is:

Does the learner retain enough internal structure to use the external object intelligently, detect obvious failure and reconstruct the required reasoning when appropriate?

For example, an engineer does not need to memorise every material constant.

But an engineer who cannot recognise an impossible order of magnitude is vulnerable even with a perfect database.

Similarly, a student using a symbolic algebra system need not perform every enormous expansion manually.

But if the learner cannot recognise an illegal transformation or impossible graph, tool use becomes dependence rather than augmentation.

13. Verification becomes more valuable as tools become more powerful

This leads to a paradox.

The better our mathematical tools become, the less valuable some routine manual operations may become.

Yet verification capability becomes more valuable, not less.

Why?

Because powerful tools increase the amount of output one person can produce.

That means errors can also propagate faster.

A learner able to produce ten calculations manually may make ten errors.

A system able to generate ten thousand calculations can distribute an incorrect assumption through all ten thousand.

So mathematical education in an AI-rich environment cannot consist only of teaching students how to obtain answers.

It increasingly has to develop the capacity to ask:

Does this make sense?
What assumptions entered the model?
Which conditions are required?
What happens at the boundary?
Is the sign plausible?
Are the units consistent?
Can another method check it?
What would falsify this result?

That is mathematical verification.

And it is one reason the BaseFloor remains important even when machines become extraordinarily capable.

14. The frontier can actually become larger once the base is protected

BaseFloor protection may initially sound conservative.

It is not.

Its purpose is to make more ambitious exploration safe.

If a student has stable symbolic control, we can let technology handle large computational searches.

If the student can interpret graphs, we can explore dynamic simulations.

If the student understands what differentiation means, we can examine numerical and symbolic derivatives computationally.

If the student understands proof and counterexample, AI can become an adversarial mathematical partner rather than an answer vending machine.

The stronger the stable base, the more aggressively the frontier can expand.

That gives us the deeper relationship:

BaseFloor does not compete with frontier learning.

BaseFloor increases the frontier that can be used intelligently.

15. A practical P3 → P4 mathematics corridor

For many topics, a useful sequence looks like this.

Stage 1 — Establish the idea

What does the mathematical object mean?

Stage 2 — Model a valid route

Study correct examples and decision points.

Stage 3 — Complete with support

Work through partially supported examples.

Stage 4 — Execute independently

Solve matched questions without the route being supplied.

Stage 5 — Delay

Return later.

Stage 6 — Vary

Change wording, representation or parameters.

Stage 7 — Mix

Require selection among several possible methods.

Stage 8 — Verify

Check through substitution, estimation, inverse operations, graphs, units, domains or alternative methods as appropriate.

At this point the operation is approaching stable P3 functioning.

Then:

Stage 9 — Extend

Introduce unfamiliar combinations, modelling, proof, computational exploration or AI.

Stage 10 — Switch the support off again

Check whether the core mathematical operation survived the extension.

Then continue.

This is not a rigid universal teaching sequence.

Receiver state matters.

A strong learner may move through the earlier stages quickly.

A fragile learner may require more guided work.

Different topics create different dependency structures.

But the controlling distinction remains useful:

Do not infer independent learning merely because sophisticated work has been produced.

16. What parents can look for

For parents, the distinction between P3 and P4 provides a calmer alternative to asking only:

“Is my child doing difficult questions?”

17. What students can ask themselves

Students can perform the same audit.

After studying a topic, ask:

Can I start the question without seeing an example?

Can I explain the first move?

Can I solve a new version tomorrow?

Can I tell when this method should not be used?

Can I detect a suspicious result?

Could I do this if the AI disappeared?

If the answer to some of these questions is no, that is useful information.

It is not a verdict on mathematical ability.

It simply identifies the next part of the system that still needs to become independent.

18. The civilisation-level implication

The same principle scales beyond one mathematics student, but it must be transported carefully.

A civilisation becomes more capable partly by creating external systems that perform tasks people once performed themselves.

Writing externalised memory.

Calculators externalised arithmetic.

Computers externalised enormous quantities of computation.

AI increasingly externalises parts of search, representation, explanation and reasoning.

Externalisation is not automatically decline.

It is one of the central mechanisms of civilisation.

But externalisation creates a new control question:

Which internal capabilities must remain distributed through the population so that society can understand, verify, repair and regenerate the external systems on which it depends?

That is the civilisation version of the BaseFloor problem.

If tools become more powerful while human verification and reconstruction capacity collapses, apparent capability can increase while system fragility also increases.

If tools grow while humans retain enough knowledge to direct, audit, question and rebuild them, civilisation has achieved genuine augmentation.

This is why mathematics matters beyond calculation.

It is part of the capability system through which increasingly powerful models of reality remain inspectable.

The direct answer

The goal of advanced mathematics education should not be to keep every student permanently operating without tools.

Nor should it be to move students toward harder and harder output regardless of how that output is produced.

The better objective is:

Build a stable independent mathematical base, then use increasingly powerful tools and problems to extend what the learner can explore—while periodically verifying that the underlying capability still belongs to the learner.

That is the relationship between BaseFloor, P3 and P4.

BaseFloor protects the prerequisites.

P3 establishes dependable independent operation.

P4 expands the frontier.

And the critical rule connecting them is:

Frontier capability must not be purchased by quietly destroying the foundation required to understand and verify it.

In the age of AI, this becomes one of the most important distinctions mathematics education can make.

Because the future student will not be measured merely by how much mathematics they can calculate alone.

They will increasingly be measured by something harder:

what they can understand, direct, verify, transfer and reconstruct while working with machines that can calculate far more than they ever could.

Research Boundary

BaseFloor, P3 and P4 are architectural constructs used by the MathematicsOS/CivilisationOS project. They are not experimentally established universal stages of mathematical development.

The research cited here supports component mechanisms—including context-sensitive transfer, the distinction between supported and unaided performance, and the importance of pedagogical design in AI-assisted learning. The mapping of these findings into the BaseFloor/P3/P4 control architecture is a project inference and should continue to be tested against learner outcomes.

Research checked

10 August 2026

The underlying master requires frontier articles to distinguish research findings from architectural synthesis and to treat AI-supported correctness separately from independent capability. Its AI-off protocol requires explanation, fresh unaided performance, representation transfer, independent verification and delayed testing before assisted work is written back as stable capability.

MATHCIV-006 | Mathematics + Civilisation Series | BaseFloor, P3 and P4: How to Advance Without Hollowing Out Foundations