A generative definition

At a broad level, people carry models that compress experience and generate expectations about the world. These models need not be explicit, globally coherent, or true. They only need to work well enough in the situations that call them forward.

Learning changes that generative machinery. The change may affect an answer, but it should also affect which relationships become visible, which conjectures feel plausible, which contradictions can be detected, and which new problems can even be conceived. Knowledge has causal properties because it changes what its holder can do next.

Boundary: The machine-learning vocabulary in this thesis is an analogy about inference, constraints, and model change. It is not an anatomical claim about the brain.

The transfer test

Suppose external support helps a learner solve a problem. The observed success belongs to the learner-plus-support system, not necessarily to the learner alone. Stronger evidence arrives later, when the changed learner can solve a sufficiently different problem that depends on the same knowledge, without the support and without being told that the knowledge is relevant.

This is why success on a familiar exercise is weak evidence. The context may announce the procedure, preserve a recently supplied explanation, or route the learner toward the answer. A novel problem tests whether the learner can now do that routing for themselves.

The prediction gap

A model earns pressure to change when what it predicts and what reality permits diverge. This difference is the prediction gap. Learning is not the mechanical minimization of every surprise, but the attempt to build better explanations when an existing one can no longer survive criticism.

The learner never needs to begin with truth. They need a model that can be made to produce consequences, a way to compare those consequences with something outside the model, and enough freedom to create a better conjecture. That loop is developed across Problems, Conjectures, Constraints, and Constructionism.

What counts as evidence

Evidence of learning is behavioral but not merely performative. It includes explaining from a blank page, recognizing relevance without a cue, combining ideas that were previously isolated, surviving a change in surface context, identifying a limiting case, and repairing an explanation after criticism.

No single artifact reveals a whole mind. The aim is therefore repeated generation across varied contexts, with support added and removed deliberately. We are not trying to certify that the learner once produced the right answer. We are trying to observe a model becoming more capable of correcting itself.

Knowledge as causal information

Calling knowledge information is only useful if information is understood causally. A sentence stored in memory is not inert when it changes which features of a situation become salient, which moves appear possible, or which consequences can be anticipated. Knowledge is information embodied in a system such that the system behaves differently because of it. The same written proposition may therefore be knowledge for one person, a recognizable phrase for another, and noise for a third. Its educational significance is not exhausted by whether it can be repeated.

This is why a model is more than a list of beliefs. It is a generative compression: a relatively small structure that can produce judgments about cases never explicitly stored. Newton's laws, a grammatical intuition, or a working model of another person's motives each compress many possible situations into relations that support prediction and explanation. A model may contain tacit procedures, images, examples, and local exceptions alongside explicit propositions. Learning can alter any of these, but the decisive change is functional: the compression now generates a different range of thought and action.

The phrase model change should not imply that an old model is simply erased and replaced with a final true one. Learners often add exceptions, reorganize relations, change which model is routed by a context, or construct a better explanation that subsumes an older one within a limited domain. Knowledge grows through corrigible improvements. The relevant contrast is not falsehood versus certainty but a model that cannot yet answer a problem versus one that explains more while exposing itself to further criticism.

Performance underdetermines learning

A visible answer is produced by an entire situation. The learner contributes prior knowledge, attention, habits, and current conjectures; the environment contributes wording, examples, social cues, tools, time, and feedback. Because many combinations of those causes can yield the same output, success under one arrangement cannot identify which internal capability produced it. A student may solve an equation by understanding invariance, imitating a recently demonstrated sequence, pattern-matching the exercise type, or following an interface that prevents illegal moves. The paper records one answer while concealing several possible generating systems.

Failure is equally ambiguous. It can indicate a missing concept, an inaccessible route to available knowledge, an overloaded representation, a mistaken local model, or a problem whose language prevents the learner from seeing what they already understand elsewhere. A serious theory of learning therefore refuses to read model state directly from a score. It changes the conditions, observes what kind of support unlocks progress, asks the learner to generate an artifact, and then removes that support to see what persists.

This distinction also explains why fluency can rise while understanding remains stationary. Repetition can make a particular path cheap and fast without making the underlying relation available outside that path. Fluency is valuable when the fluent component participates in wider reasoning, but speed alone is not the definition of learning. The question is always what new behavior the learner's changed model can generate when the original prompt, sequence, or helper is no longer doing the routing.

An evidence hierarchy

Evidence for model change comes in degrees. Recognition is weaker than recall because the answer remains present as a constraint. Recall in the original wording is weaker than reconstruction in a new representation. Reproducing a demonstrated procedure is weaker than selecting it when the problem does not name the method. Solving a near-transfer exercise is weaker than explaining why the relation survives a changed surface context. Generating a novel consequence, identifying a counterexample, or integrating the idea with an apparently conflicting model is stronger still because more of the relevant structure must be produced internally.

This hierarchy is not a universal ladder on which every lesson must climb. Some knowledge is appropriately tested through rapid recognition; some motor or perceptual learning is revealed by skilled performance; some explanations require tools that experts also use. The principle is conditional: evidence should match the capability being claimed. If the claim is independent explanatory knowledge, the assessment must remove the external structures that would otherwise supply the explanation. If the claim is intelligent tool use, the tool can remain, but the learner must still choose, direct, and criticize its contribution.

Repeated evidence matters because a model is distributed across contexts. One successful transfer may be luck or a surface analogy. A trajectory of constructions is more revealing: the learner needs fewer prompts, detects deeper errors, connects previously separate ideas, and rebuilds more precisely after criticism. The evidence is not that errors disappear. It is that the learner becomes better at producing errors worth criticizing and at using criticism to reorganize the model that produced them.

Learning changes the frontier

The deepest evidence of learning is sometimes not a solved problem but a newly visible one. Before learning, an inconsistency may pass unnoticed because the learner lacks the concepts required to formulate it. After a model changes, the same situation can become surprising. New knowledge increases explanatory reach while creating a sharper boundary between what is and is not yet explained. The learner is able to ask a better question because the previous question has become part of the machinery with which they now think.

This makes learning recursively productive. A model generates consequences; those consequences meet reality; discrepancies become Problems; problems provoke Conjectures; criticism selects among them; and the resulting model exposes a new frontier. Education should therefore not optimize for a terminal state in which the learner has no questions. It should cultivate systems that can find and pursue increasingly consequential problems without waiting for an institution to specify the next exercise.

Autonomy is not the absence of dependence on other people or accumulated knowledge. No learner independently recreates civilization. It is the capacity to use inherited explanations as material for further construction: to know when an answer is only borrowed, when a tool is silently performing the inference, when a contradiction deserves attention, and what kind of artifact could expose the next conjecture to criticism. That self-correcting capacity is the durable form of model change this thesis calls learning.

Start with the objective

If the objective of schooling is to educate, and education exists to facilitate learning, then the design of school should follow from a theory of learning. The current order is usually reversed. Institutions begin with schedules, subjects, standards, assessments, and content, then ask learners to adapt to the machinery.

A first-principles design begins with a different question: what must happen for a learner's model to become more capable of explaining and acting in the world? Every educational component then has to justify itself by the role it plays in that change.

Information is an input

Information can be transmitted. Understanding cannot. A teacher can transmit an explanation, equation, diagram, example, or procedure. The learner still has to interpret that representation through a different history of experiences, associations, conjectures, and contradictions.

This does not make instruction useless. Information from other people is one of civilization's great accelerators. It means instruction should be judged by the construction it enables, not by the elegance of the representation alone.

What school gets right

School is good at reducing an otherwise enormous search space. Subjects, sequences, teachers, examples, and problem sets can place a learner near knowledge they would be unlikely to discover unaided. In the language of this thesis, school supplies Constraints.

The failure begins when supplying constraints becomes the entire theory of education. A path can be narrowed until the learner has almost nothing left to search, generate, or criticize. They reach the desired conclusion while the institution mistakes its own guidance for the learner's understanding.

The design question

The design target is not maximum freedom or maximum instruction. It is enough constraint to make construction possible, but not enough to make construction unnecessary. The right amount varies with the learner, the problem, and the moment.

A good environment progressively gets itself out of the loop. It helps a learner do something they cannot yet do alone, then removes support until the capability belongs to the learner. The transition from augmented performance to autonomy is the actual product.

Education is not schooling

Schooling is an institutional arrangement; education is a function an arrangement may or may not perform. A school coordinates time, people, credentials, custody, socialization, curriculum, and access to cultural knowledge. Those functions can be valuable without being identical to learning. The distinction matters because institutional outputs are easy to count and model change is difficult to observe. Attendance, content coverage, compliant work, completed courses, and examination scores can gradually become proxies that replace the objective they were meant to indicate.

Education also happens outside school whenever an environment makes reconstruction possible: a workshop in which materials resist a design, a conversation that exposes a contradiction, a book that creates a problem the reader cannot ignore, or a community whose standards make weak explanations visible. Conversely, a beautifully organized lesson can fail educationally if the learner only follows a path whose important choices have already been made. The institutional label tells us almost nothing until we inspect what the learner must generate and what remains after the situation ends.

Separating the concepts prevents a false reform strategy. If education is assumed to be whatever schools currently do, improvement means making delivery faster, clearer, more personalized, or more measurable. If education is the deliberate design of conditions for model change, every existing component becomes contestable. The timetable, subject boundary, lecture, age cohort, grade, and classroom are means with owners and tradeoffs, not natural laws of learning.

The limits of instructionism

Instructionism treats the central educational problem as finding the correct representation of expert knowledge and delivering it accurately. This solves a real problem: civilization contains explanations no learner could efficiently recreate, and poor representations impose needless difficulty. But it mistakes successful encoding for successful reconstruction. The teacher compresses a rich internal model into words, diagrams, equations, demonstrations, and examples. The learner receives those external forms, not the causal network from which the teacher produced them.

The learner must interpret the representation through a different model. Familiar words may attach to different examples; an equation may be manipulated without its invariants being understood; an analogy may preserve a surface similarity while importing the wrong causal structure. Increasing the clarity of the transmission can reduce friction yet also increase the feeling of understanding before the learner has generated anything. The resulting Comprehension illusion is not an argument against explanation. It is an argument against treating explanation as the completed unit of learning.

A representation earns educational value through what follows it. Does the learner predict before seeing the outcome, reconstruct the relation in another form, choose where it applies, use it with another idea, or discover a consequence the explanation did not explicitly state? Instruction belongs inside a larger loop of construction and criticism. It can supply indispensable Constraints, but it cannot substitute for the learner reorganizing their own generative structure.

What an environment actually controls

An educational environment cannot select the exact internal change a learner will make. It can control the distribution of encounters: which Problems appear, which tools are available, when explanations arrive, what kinds of artifacts are required, whose criticism matters, how costly an error is, and whether support remains present during assessment. It also shapes attention through norms. A classroom can make guessing socially dangerous or make revision ordinary; it can reward rapid correctness or reward the discovery of a hidden assumption.

Teachers are therefore more than transmitters and less than programmers of minds. They diagnose the current construction through its consequences, select constraints that make a better conjecture reachable, protect productive confusion from premature closure, and remove assistance when it has done its work. Tools play a similar double role. A simulation can create consequences unavailable in a textbook, while an interface can also hide the very relation the learner needs to construct. AI can widen the range of feedback and representations, while silently completing the intellectual act being assessed.

Institutions act at another scale. They determine how much iteration is affordable, whether learners can follow a problem across subject boundaries, whether teachers have time to inspect reasoning, and whether credentials reward durable capability or short-term compliance. None of these elements causes learning in a linear way. Together they define a landscape in which some model changes become probable, others become unnecessarily expensive, and many never become thinkable.

A first-principles environment

A first-principles design starts from the smallest defensible learning loop: a learner has a model; a problem makes its limit matter; the learner constructs a conjecture or artifact; consequences criticize it; the learner rebuilds; and the environment later tests what can be regenerated without the original support. Curriculum then becomes a sequence of increasingly powerful problem spaces, not merely an ordered inventory of statements. Assessment becomes an instrument for locating dependence and change, not only ranking completed outputs.

One-to-one attention is valuable when it makes this loop observable. A construction-aware tutor can notice which representation the learner produced, distinguish a missing prerequisite from a routing failure, and alter the next constraint accordingly. The ideal is not an infinitely patient lecturer generating endless personalized explanations. It is an environment with enough resolution to see the learner's conjectures and enough restraint to avoid replacing them. This is the mathetic interpretation of personalization.

The environment succeeds by becoming progressively less causally necessary. Early on, it may choose the problem, supply examples, model criticism, and structure the artifact. Later it should withdraw each function the learner can assume. The strongest educational outcome is not dependence on an exceptionally effective system. It is a person who has appropriated the system's useful functions: finding problems, seeking information, building tests, inviting criticism, and continuing the cycle without permission.

Why bad models persist

A child's expectation that moving objects naturally stop is locally reasonable in a world full of friction. The model explains everyday experience well enough. It has no reason to include inertia until a situation such as skating on ice makes the old explanation visibly incomplete.

Wrong models are often not irrational. They are compressed explanations adapted to a limited range. Simply declaring them wrong adds information but may not produce the problem that would make reconstruction necessary.

Problem selection

A productive problem sits beyond what the learner can already explain but within reach of a search that available knowledge and support can constrain. Too easy and no reconstruction is required. Too remote and the learner has no meaningful way to conjecture.

This is why frontiers matter at both individual and civilizational scales. The edge of a model is where its assumptions are least tested, where contradictions accumulate, and where new knowledge can alter the space of future problems.

Problems and agency

A learner who owns a problem searches differently from one who has merely been assigned an exercise. Ownership does not require that every topic be self-chosen. It requires that the learner can perceive what is at stake and why the present model is insufficient.

The educational move is therefore not to hide all explanations. It is to preserve a genuine problem long enough for the learner to form a prediction, make a guess, and experience the need for a better model before the solution closes the space.

A problem is not an exercise

An exercise usually begins after the important intellectual decisions have been made. Its location in a chapter identifies the relevant concept, its wording implies the permitted variables, and its similarity to a worked example suggests the procedure. It may provide valuable fluency, but the learner's task is often to execute a route rather than determine what kind of situation they face. A problem begins earlier. The learner must decide what is surprising, which facts constrain an explanation, and what would count as a resolution.

The same prompt can be an exercise for one person and a problem for another. For the expert, a differential equation may instantiate a familiar class; for the novice, even identifying the changing quantities may require construction. Conversely, a child's question about why the moon appears to follow the car can be a genuine explanatory problem despite requiring no formal notation. Problem status is a relation among a demand, a model, and a context—not a property printed on a worksheet.

This also means difficulty is not the same as problem quality. A long calculation can be difficult while leaving the model untouched; a simple observation can destabilize an entire explanation. Productive difficulty is located at the point where the learner must reorganize relations, not wherever the task consumes the most time. Educational design should remove accidental difficulty when it obscures the target while preserving the incompatibility that gives new knowledge a reason to exist.

Locally useful models

Models persist because they solve some class of problems. The intuition that motion requires a continuing push works in ordinary environments where friction is nearly universal. The story of Santa Claus organizes scattered evidence for a young child because testimony, rituals, presents, and social authority all point in the same direction. Calling either model irrational misses its local economy. Each compresses the available evidence without requiring concepts the learner has not yet constructed.

A new explanation becomes necessary when the old one is forced beyond the range in which it works. Ice reduces friction enough for inertia to become a lived discrepancy. Conflicting logistical details or social knowledge can make the Santa model increasingly expensive to preserve. The educational task is not merely to announce the replacement. It is to design an encounter in which the learner's current story produces a consequence that cannot be reconciled cheaply, thereby converting external information into an internally owned problem.

Local usefulness also explains why misconceptions survive correction. The corrected answer may occupy a school context while the older explanation continues to organize everyday experience. Unless an artifact requires both models to constrain the same consequence, they need not compete. A problem is powerful when it crosses those contextual boundaries and makes the cost of incoherence visible. This is why Context and coherence is inseparable from problem design.

Problems make knowledge compound

Knowledge does more than supply answers; it changes the distribution of possible problems. Before a learner understands natural selection, variation among organisms may appear as disconnected description. Afterward, the distribution of traits, the cost of an adaptation, and the mismatch between organism and environment become explanatory questions. A new model compresses previous observations and simultaneously makes finer discrepancies visible. What once looked like noise becomes structured evidence against or for competing conjectures.

This compounding is why foundations matter without implying a rigid universal sequence. Prior explanations constrain search, allowing the learner to reject vast regions of possibility without exploring them. But what counts as foundational depends partly on the problems being pursued. Programming can make algebraic structure concrete; a historical controversy can create the need for statistical reasoning; building a machine can create the need for geometry. Knowledge and problems form a reciprocal graph rather than a one-way staircase.

An educational system that only assigns problems with known procedures captures little of this compounding. It gives the learner destinations without transferring the capacity to expand the map. A stronger sequence alternates resolution with frontier creation: each explanation solves something that mattered, then reveals a class of questions the learner can now formulate and investigate with greater independence.

Individual and civilizational frontiers

For an individual learner, a problem can be old to civilization and still be genuinely new to the model doing the work. The learner need not believe they are the first person to solve it. What matters is that the explanatory route has not been supplied so completely that the learner's own search becomes ceremonial. Reconstructing an inherited insight is different from historically discovering it, but both require a conjecture to be generated, constrained, and criticized inside a model.

At a civilizational frontier, no teacher possesses the final route. The Wright brothers could draw on accumulated knowledge of engines, structures, gliding, and control, yet controlled powered flight remained a problem because those pieces had not been organized into a working explanatory system. Their progress depended on artifacts that made errors measurable—gliders, wind-tunnel data, control surfaces—and on rejecting authoritative assumptions when consequences failed. The frontier magnifies the same loop present in a learner's smaller reconstruction.

Education should prepare learners for both frontiers. If all problems arrive with chapter labels, approved methods, and known answer forms, the learner becomes skilled at navigating institutional certainty. Real work often begins when the category is unclear, authorities disagree, evidence is incomplete, and constructing the test is part of constructing the explanation. Agency means being able to remain inside that uncertainty without confusing unconstrained opinion with a conjecture exposed to reality.

Explanations are stories with consequences

If a learner says that a force gives an object motion that the object gradually uses up, they have told a causal story. Everyday observations may appear to support it. The story becomes educationally useful when it is pushed into a situation where it predicts something nature refuses to do.

Reality does not hand over an explanation. It constrains the space of explanations by ruling out stories whose consequences fail. This is why observation without conjecture cannot do the whole job: data has to criticize some proposed account of what is happening.

Search is never random

The space of possible guesses is effectively unbounded, but a learner never searches it from nowhere. Existing knowledge, memories, analogies, goals, tools, and the problem itself make some ideas thinkable and others invisible.

Knowledge compounds partly because it changes this search. A better explanation does not only solve the current problem; it removes dead regions, opens new questions, and alters which future conjectures can be generated at all.

Commit before explanation

A prediction is a compact artifact of a conjecture. Asking for it before explanation prevents the learner from retrospectively adopting the correct story without revealing what their own model would have generated.

Commitment should not become humiliation. Its purpose is measurement. The prediction creates a stable object that can be compared with consequence, discussed, revised, and revisited after support is removed.

Observation does not contain its explanation

No finite collection of observations uniquely specifies the story that produced it. The sun's apparent motion can be described by a moving sun, a rotating Earth, or more elaborate arrangements engineered to preserve the same appearances. A falling object supplies positions over time, not the concept of gravity. Data constrains explanation only after an explanatory structure has generated consequences with which the data can disagree. Theories are therefore not summaries waiting inside observations to be extracted by sufficient attention.

This does not reduce knowledge to arbitrary storytelling. Conjectures differ in reach, precision, internal coherence, compatibility with other explanations, and resistance to criticism. Reality can decisively eliminate possibilities even when it cannot dictate the next one. The creative act supplies candidate structure; criticism removes structures that fail. Learning requires both. Pure reception omits creation, while unconstrained expression omits the selective pressure that makes better knowledge possible.

Educationally, telling learners to 'discover the pattern' can conceal this structure. A pattern is noticed through concepts the learner already has, and many patterns fit the same cases. A more honest task asks for a candidate explanation, requires the learner to state what else would follow if it were true, and then seeks a case that distinguishes it from a rival. The object is not to reproduce the teacher's path of discovery but to participate in the logic by which explanations become criticizable.

Conjecture as constrained search

A learner's possible conjectures can be imagined as a latent space: an enormous set of explanations, analogies, procedures, and combinations that could in principle be generated. The metaphor is useful because search rarely proceeds by enumerating explicit options. A word, diagram, prior problem, bodily intuition, or social cue changes which regions become reachable. This is an analogy about generation and routing, not a claim that human thought implements a particular machine-learning architecture.

Prior knowledge is powerful partly because it deletes search. Conservation laws rule out entire families of perpetual-motion explanations. Evolutionary theory makes purposeful design unnecessary as the default biological account. A concept such as opportunity cost reveals consequences that remain invisible inside a purely monetary frame. Expert speed often reflects this transformed space: not faster traversal of all possibilities, but a model that does not generate most novice errors as live candidates.

Constraints supplied by a teacher or tool can temporarily create the same effect from outside. A hint may direct attention to a relevant invariant; a template may prevent malformed solutions; an AI may propose the missing analogy. That can make progress possible, yet the learner has learned the constraint only when a changed internal model later removes or deprioritizes the same impossible regions without the helper. Thus conjecture and Constraints must be analyzed together.

Guessing needs criticism

Inviting guesses is not a celebration of whatever first comes to mind. A conjecture becomes intellectually serious when the learner takes responsibility for its consequences. What would we expect to observe? Which cases should become impossible? What would distinguish this story from another that fits the same initial facts? These questions turn imagination into an object that can lose. Without that possibility, the activity remains expression rather than explanation.

Criticism can come from several directions. Reality may refuse a prediction; the conjecture may contradict itself; it may conflict with a stronger explanation elsewhere; a peer may expose an unstated assumption; or an artifact may fail under a limiting case. The best criticism is selective: it locates why a model fails without replacing the entire construction for the learner. A correction that supplies every missing relation can produce agreement while removing the need to rebuild.

A healthy culture separates being wrong from being careless. Wrong conjectures are necessary because no search begins with guaranteed truth. Carelessness means protecting a conjecture from consequences, changing it after the fact so it cannot fail, or borrowing an explanation without being able to regenerate its logic. The classroom should make bold, explicit guesses cheap and uncriticizable vagueness expensive.

Commitment is for revision, not identity

Committing to a prediction before receiving the explanation establishes a temporal boundary between what the learner's model generated and what the environment supplied. That boundary is essential for diagnosis. If the answer is revealed first, people are remarkably able to reinterpret their earlier uncertainty as near-understanding. A written prediction, diagram, or causal account preserves the prior construction so the learner can compare models rather than merely experience the fluency of the new one.

The commitment must remain provisional. If classrooms turn predictions into public tests of intelligence, learners rationally minimize risk: they wait for cues, imitate authority, and produce claims too vague to criticize. The desired norm is strong attachment to the process of explanation and weak attachment to any particular conjecture. Revision should be evidence of successful criticism, not a social confession.

The most informative sequence is therefore predict, explain, observe, criticize, rebuild, and later regenerate. Each stage produces an artifact with a different diagnostic role. The initial prediction reveals routing; the explanation reveals causal structure; the comparison reveals the precise gap; the reconstruction reveals how the learner responds to criticism; and the delayed regeneration tests whether the improved constraint has become part of the learner rather than remaining in recent context.

A continuum, not a binary

The choice is not discovery learning versus direct instruction. Every act of learning occurs inside constraints. Even a supposedly open project has materials, goals, prior knowledge, physical laws, social expectations, and available time.

The useful question is how much of the search each constraint performs. A small hint may unlock knowledge the learner already has but failed to route. A worked solution may supply the concepts, assumptions, and sequence so completely that only recognition remains.

Support is diagnostic

Two learners can fail the same problem while needing radically different help. One may need a change of representation; another may lack a prerequisite; a third may possess the relevant knowledge but only in a context that was not activated.

The amount and type of constraint required therefore reveals more than the final answer. Add support in small, legible increments and observe how the learner's reasoning changes. Then remove it and test whether the new capability remains.

Add, observe, remove

Good support is adaptive and temporary. It reduces search enough for the learner to generate a next move, then fades so that the learner must carry more of the structure. The goal is not perpetual successful performance under ideal guidance.

This makes constraint removal as important as constraint delivery. If an educational system never tests the learner after support disappears, it cannot distinguish a capability that belongs to the learner from one rented from the environment.

A taxonomy of constraints

The first constraint is the problem itself. A demand rules out explanations that do not address it. Reality adds another boundary by refusing consequences the world does not permit. Prior knowledge supplies internal constraints: a learner who understands conservation, causation, or proportionality does not need to test every conceivable story. Language, notation, diagrams, examples, tools, and materials shape which features can be represented and manipulated. Social constraints—authority, peer judgment, grades, and norms about error—shape which conjectures a learner is willing to expose.

Teachers and curricula add deliberate constraints. A sequence can ensure that a powerful prerequisite is available before a problem requires it; a hint can name a relevant distinction; a worked example can demonstrate a search strategy; a question can force an assumption into view. AI can generate all of these dynamically and at high resolution. Yet the source does not determine the effect. The same example can activate an existing structure for one learner, replace the construction for another, and overload a third with irrelevant detail.

Constraints differ in kind as well as quantity. Removing choices from an interface is not equivalent to giving a conceptual hint. Naming the theorem is not equivalent to drawing attention to an invariant. Supplying the next step is not equivalent to asking for a prediction. A useful analysis asks exactly which uncertainty a constraint removes and whether that uncertainty was incidental friction, a missing prerequisite, or the intellectual work through which the target model would have been constructed.

Too few and too many

With too few constraints, the search space becomes arbitrary. A novice asked to derive a scientific theory from raw observation lacks the concepts required even to describe the relevant regularities. Failure produces little information because almost any missing structure could explain it. The learner may resort to blind trial, superficial pattern matching, or waiting for approval cues. Calling this freedom does not make it educational; agency requires a model capable of making meaningful choices.

With too many constraints, the learner follows a corridor whose walls encode the solution. Fill-in-the-blank notes, color-coded procedures, leading questions, auto-completion, and stepwise hints can create nearly errorless performance. Errorless performance may be appropriate during part of training, but it removes evidence about which decisions the learner can originate. When every relevant relation remains present, recognition and compliance masquerade as generation.

The productive region resembles the zone of proximal development, interpreted here as a moving relation between model, problem, and available help. The learner cannot yet complete the construction alone, but a limited constraint makes a meaningful next conjecture possible. The zone is not a fixed level of task difficulty. It changes after every model revision and differs across representations, contexts, and kinds of support.

The support ladder as measurement

Instead of classifying a response only as correct or incorrect, an environment can vary support deliberately. Begin with the problem under minimal cueing. If the learner stalls, change the representation without supplying the relation. Then identify a relevant feature, remind them of a prior case, expose a contradiction, provide a partial structure, or finally model the full solution. The point at which construction restarts reveals what kind of dependence exists.

The shape of help matters more than a raw hint count. A learner who succeeds after the context changes may have possessed the concept but failed to route it. One who succeeds after a prerequisite is recalled may have a disconnected knowledge graph. One who can imitate only after the full sequence is modeled does not yet possess the generative structure. Two identical answers at the end of this ladder therefore encode very different next educational moves.

This diagnostic use of support requires legibility. If an AI continuously rewrites the problem, proposes steps, checks algebra, and supplies encouragement in one stream, its contribution cannot be separated from the learner's. Better systems make constraints explicit, add them in discrete increments, and record what changed. The evidence is the trajectory of generation under changing support, not merely the polished terminal artifact.

When an external constraint becomes internal

A constraint has been learned when the learner can regenerate its selective effect. At first, a teacher may ask, 'What is conserved?' Later the learner spontaneously searches for an invariant. A checklist may initially require sources of error; eventually the learner notices them while designing the experiment. A worked example may demonstrate separating a problem into cases; later the decomposition appears without the example. The content of the help has become part of how the model searches.

Fading should therefore remove not only visible aids but also routing cues. Giving the same problem without the formula sheet tests recall, but leaving it under a chapter called Conservation still announces the relevant model. Stronger fading changes the surface context, mixes problem types, delays the attempt, and asks the learner to decide what information is needed. What persists is not a memorized prompt-response pair but an internally available constraint on future conjecture.

The destination is not a learner without tools. Experts intelligently preserve external constraints: notation, reference works, software, collaborators, and instruments extend what any unaided mind can do. The distinction is control. An autonomous learner can select the tool, understand which uncertainty it removes, criticize its output, and continue when it is unavailable. A dependent learner is carried by a constraint whose intellectual contribution they cannot yet identify.

The augmented system

When a learner succeeds with help, we have observed the coupled system of learner and support. Nothing in that result proves that the learner can later reproduce the result alone or recognize where the same knowledge matters.

The analogy to AI context is useful but limited. Adding a skill file to a model's context can transform its output without changing its underlying parameters. In the same way, surrounding a person with the right explanation can improve current inference without becoming part of how they independently think.

Boundary: The analogy distinguishes external augmentation from durable change. It does not claim that human learning is gradient descent or that brains are language models.

The comprehension illusion

Clear explanations produce a powerful feeling of recognition. Each step makes sense while it is present, so the learner mistakes the fluency of the explanation for the availability of the idea inside their own model.

Generative AI amplifies this illusion because it can instantly rewrite, personalize, simplify, and continue. A learner can accumulate the sensation of understanding across a curriculum while never being required to generate, combine, or criticize the underlying ideas.

Stronger evidence of change

Wait until the supplied context is no longer fresh. Change the surface features of the problem. Remove the cue that announces which concept applies. Ask for an explanation, prediction, or derivation rather than recognition. Combine familiar ideas in an unfamiliar relationship.

These moves do not make assessment arbitrarily difficult. They test the claim education actually cares about: that knowledge once supplied from outside can now be generated and routed by the learner when it is relevant.

The formal distinction

Let $M$ denote the learner's present model, $S$ an external scaffold, $x$ a task, and $y$ a successful output. Observing $(M + S)(x) \rightarrow y$ establishes only that the coupled system can produce $y$. The educational claim we usually want is stronger: some process has produced $M^{\prime}$ such that $M^{\prime}(x_j) \rightarrow y_j$ across a relevant family of cases, including cases where $S$ is absent and the applicability of the knowledge is not announced. The prime marks a durable change in the learner, not merely a better result.

The formalism is intentionally minimal. It does not tell us how a human model is represented, how durable a change must be, or how far transfer should extend. Its purpose is to prevent a causal inference error. If adding S and adding learning can both improve the same observed output, the output under S cannot distinguish them. We need interventions: remove the scaffold, delay the attempt, alter the context, or ask for a construction whose critical relations were previously supplied.

Even $M^{\prime}(x) \rightarrow y$ is insufficient if $x$ is identical and $y$ can be cached as a response. The family $\{x_j\}$ matters because learning claims always imply some invariance beneath changing surface conditions. The appropriate variation depends on the knowledge. A pianist need not transfer a fingering to chemistry, but should retain control when tempo or phrasing changes. A principle of proportionality should remain available when the quantities and notation differ.

Inference and training

The difference resembles the distinction between changing an AI system's context and changing its parameters. A prompt can supply definitions, examples, procedures, and goals that reorganize current inference. Remove the prompt and the base system may return to its former behavior. Training aims to alter the system that will generate later outputs. In the educational analogy, scaffolding behaves like temporary context while learning changes the system that will act later.

The analogy is valuable because it makes augmentation visible. A student with a live AI assistant may form a highly capable joint system: the human identifies a goal, the AI retrieves and proposes, and the human selects or edits. That joint capability can be genuinely useful without belonging entirely to either component. Trouble begins only when an institution attributes the coupled system's output to an unaided learner or claims that repeated exposure to the assistant has necessarily changed the learner.

The boundary must remain explicit. Human learning is not parameter updating in a language model, recent context is not a literal context window, and social, embodied, emotional, and developmental processes cannot be reduced to the analogy. It is a causal abstraction: changing inputs around a system and changing the system are different interventions, even when both improve the next answer.

Assessment inside the scaffold

Many assessments preserve the same cues used during instruction. Problems remain grouped by chapter; notation matches examples; the required procedure is recent; partial-credit prompts reveal the sequence; and formula sheets identify the relevant variables. Such assessments can measure supported execution, which may be useful, but they cannot by themselves justify claims about independent selection, explanation, or transfer.

Removing every support is not automatically more authentic. Experts use calculators, documentation, collaborators, and AI. If the target capability includes those tools, assessment should preserve them while shifting the burden of judgment: the learner chooses when to use the tool, formulates the query, checks assumptions, detects an implausible answer, and integrates the result into a larger explanation. The aim is not ritualized unaided performance; it is correct attribution of which part of the capability resides where.

A robust assessment matrix varies delay, cueing, context, and tool access independently. Immediate performance with help shows reachable capability. Immediate performance without help shows current independent fluency. Delayed reconstruction tests persistence. Uncued transfer tests routing. Critique of a plausible wrong answer tests the structure of explanation. Together these observations estimate model change more honestly than a single score.

AI intensifies the attribution problem

Generative AI does not merely provide information. It can infer intent, select examples, decompose tasks, maintain context, repair syntax, simulate criticism, and generate the final artifact. These are precisely the functions through which a learner's model would otherwise become visible. The better the assistant becomes at anticipating the next intellectual move, the easier it is for the learner to remain inside a stream of local agreement without constructing the relationships that make the answer independently generative.

The subjective signal is misleading. Personalized prose feels unusually clear because the model can remove each point of friction as it appears. Yet friction is heterogeneous: some is irrelevant wording, some is a missing prerequisite, and some is the prediction gap that should drive reconstruction. An assistant optimized only for smooth continuation deletes all three. It can therefore maximize comprehension-as-feeling while minimizing evidence of comprehension-as-capability.

A learning-oriented AI should sometimes refuse to complete the loop. It can ask for a prediction before supplying the explanation, expose a counterexample instead of announcing the error, label which hints have been consumed, invite the learner to compare artifacts, and later test regeneration with fewer cues. Its success metric should not be how consistently the human-plus-AI system produces polished work. It should be how much of the valuable work the human can eventually originate, direct, and criticize.

Not a project-based-learning feature

Constructionism is often reduced to assigning projects. That misses the mechanism. A project can be so prescribed that it exposes almost nothing, while a carefully chosen explanation or unfamiliar test problem can require substantial construction.

The relevant distinction is how much of the learner's model an activity forces into the open and how predetermined the output is. Rich artifacts make more relationships interact, creating more chances for incoherence to become visible.

Information can be given; structure cannot

Learners do not need to rediscover the history of civilization. Existing explanations should constrain their search. What cannot be copied into a learner is the internal structure that makes those explanations meaningful and generative in a new context.

Construction is the work in the middle: what the learner does with received information so that it becomes connected to prior knowledge, participates in new explanations, and changes future inference.

The destination is autonomy

An environment can first help a learner build artifacts they could not yet produce alone. Across repeated attempts, it should remove the external constraints until the learner can generate, criticize, and reconstruct explanations independently.

At that point the environment has not merely delivered an answer. It has changed who is capable of producing the next one. The aim is to get the educational system out of the loop wherever the learner can now carry it themselves.

Piaget, Papert, and Popper

Three traditions converge in the mechanism proposed here. Piaget's constructivism treats knowledge as actively organized by the knower rather than copied from the environment. Papert's constructionism adds that this internal construction becomes especially powerful when it occurs through making a public object that can be shared, inspected, and revised. Popperian epistemology adds conjecture and criticism: knowledge grows through explanatory guesses that expose themselves to failure, not through the passive accumulation of observations.

Combined, these claims yield a loop. A learner's current model generates an external construction. The construction has consequences that reality, other people, or the learner can criticize. The mismatch becomes a problem for the model that produced it. The learner then reconstructs both artifact and explanation. The artifact is not evidence added after learning; it is part of the environment through which the model becomes able to criticize itself.

This synthesis also limits romantic interpretations of making. Construction without criticism can stabilize error, and criticism without learner construction can merely substitute an authority's model. Public activity without model change is performance. The educational value arises from the full circuit: internal structure produces something determinate, that something can be wrong in a locatable way, and the learner performs the work of rebuilding.

Construction is not synonymous with projects

A project has no automatic claim to constructionism. If every step, component, and success criterion is prescribed, the learner may assemble an object while making almost no explanatory decisions. The visible product can be large while the exposed model is small. Conversely, predicting the path of a pendulum, writing a proof from minimal premises, explaining a word's structure, or debugging ten lines of code can require the learner to coordinate many relationships in a compact artifact.

The relevant dimensions are generativity, coupling, and criticizability. Generativity asks how much of the output the learner must originate. Coupling asks how many parts of the learner's knowledge must jointly constrain it. Criticizability asks whether the artifact produces consequences precise enough to reveal a failure. Scale, duration, physicality, and aesthetic polish may enrich an experience, but they do not define the learning mechanism.

This rescues both tests and projects from crude categories. An unfamiliar examination problem can be deeply constructionist if the method is not announced and the reasoning must be externalized. A semester project can be non-constructionist if a tutorial silently encodes every decision. Educational design should inspect the causal burden carried by the learner, not the genre label attached to the assignment.

From one model to another

Suppose a teacher begins with knowledge $L_A$, compresses it into a communicable artifact $C_A$, and transmits some representation $X$. The educational path is not $L_A \rightarrow L_B$. It is $L_A \rightarrow C_A \rightarrow X$, followed by $X$ entering a learner who already possesses $L_B$. The learner constructs something from $X$ and $L_B$, producing $C_B$ and, if the loop succeeds, a changed model $L_B^{\prime}$. Every arrow can lose, transform, or add structure.

This explains why the teacher's most elegant explanation may be the wrong educational object. $C_A$ is optimized by the teacher's compression; its brevity depends on relations already present in $L_A$. The learner cannot simply decompress it into the same network because $L_B$ defines a different space of associations and omissions. A good representation is therefore not merely faithful to expert knowledge. It is chosen for the construction it makes possible from this learner's present model.

Papert's example of learning the word 'flower' through its relation to flour and the idea of what is best illustrates the principle. An etymological or analogical connection is not valuable because it is the canonical way to encode the word. It is valuable when it gives a particular learner material with which to construct a memorable and generative relation. The learner's route need not mirror the teacher's route to produce knowledge with comparable explanatory power.

Construction makes criticism possible

A private sense of understanding is difficult to inspect because it can change as soon as the correct answer appears. Construction creates commitment. A program either produces behavior, an argument has inferential joints, a model generates a prediction, and a diagram encodes spatial relations. The artifact freezes enough of the learner's current organization for a discrepancy to have a location. Criticism can address a missing edge, false assumption, unsupported transition, or failed consequence rather than the vague identity of being wrong.

The richest constructions permit several critics. Reality tests whether the artifact works. Peers reveal alternative interpretations. Existing explanations supply powerful constraints. The learner can compare present and prior versions to see which relations changed. A teacher can choose the smallest intervention that reopens construction. These critics should not converge too early on correcting the surface product; the real object of interest is the model that will generate the next product.

Repeated construction changes the learner's role. Initially, the environment selects problems and supplies standards of criticism. Over time the learner learns to ask what an explanation predicts, design a test, seek a limiting case, and recognize when an artifact is too predetermined to reveal anything. Autonomy is the internalization of this critical environment: the learner becomes capable of constructing not only answers but the conditions under which their own answers can fail.

Tests and projects are not opposites

A multiple-choice item can expose almost nothing beyond selection under a narrow context. An unfamiliar exam problem that asks for a derivation and explanation can expose many interacting parts of a learner's model. Likewise, a project assembled from copied steps may expose very little.

The right axis is not test versus project. It is recognition versus generation, isolated procedure versus interacting knowledge, and predetermined path versus meaningful conjecture.

Artifacts create surfaces for criticism

A thought that remains private can move, blur, and retrospectively conform to the answer. An artifact stabilizes it. The learner and the environment can point to a particular assumption, consequence, missing relationship, or contradiction.

Criticism should target the artifact and the explanation, not the identity of the learner. Error is useful because it provides location. The goal is not to punish the wrong output but to make reconstruction more precise.

Designing a revealing artifact

A revealing artifact requires the learner to choose what matters, combine more than one idea, and produce a consequence before seeing the canonical answer. It can then be tested against reality, a counterexample, another construction, or a changed context.

• Ask for a prediction and the assumptions behind it.
• Change the context so the relevant concept is not announced.
• Require two familiar ideas to constrain the same output.
• Ask what the world would look like if the learner's explanation were true.
• After feedback, ask the learner to rebuild rather than merely acknowledge.

A taxonomy of artifacts

Predictions are small artifacts that reveal what a model expects before the outcome supplies a cue. Explanations expose causal structure and assumptions. Drawings and diagrams externalize spatial or relational organization. Derivations reveal which transformations the learner treats as legitimate. Arguments expose dependencies among claims. Programs make procedural knowledge executable. Experiments and tests embody a conjecture about what observation could distinguish competing stories. Projects coordinate many of these forms across time.

Each artifact illuminates some dimensions while hiding others. A correct prediction can be a guess; an articulate explanation can be borrowed; a functioning program can be assembled from snippets the learner cannot debug; a polished project can conceal unequal collaboration. No format is transparent access to a model. Reliability comes from triangulation: ask the learner to move between forms, explain a design decision, predict a change, diagnose a failure, or reconstruct a central relation after the original artifact is gone.

Artifacts also include the design of inquiry itself. Choosing what to measure, specifying a counterexample, writing a test case, or deciding which evidence would change one's mind externalizes higher-order knowledge about criticism. These artifacts are especially valuable because they reveal whether the learner can create a problem environment rather than merely respond within one.

Richness and predetermination

Artifact richness is the amount of a model that must participate in generating the output. A one-word response may require a single association; an explanation that coordinates mechanism, evidence, boundary conditions, and a counterexample forces several structures to interact. Richness is not identical to length. A short proof can be extremely rich if each step depends on a deep relation, while a long report can remain a sequence of disconnected summaries.

Predetermination measures how much of the artifact's structure is supplied by the environment. Templates, rubrics, starter code, sentence frames, worked examples, and AI completions can all make construction reachable. They also occupy decisions the learner might otherwise make. The design question is not whether to use them but which uncertainty they remove. A template that handles formatting may reveal conceptual reasoning; a template that names every conceptual category may conceal it.

The two dimensions interact. A rich project can be heavily predetermined, and an open prompt can demand only a shallow opinion. The strongest artifact for a given moment carries enough structure to focus criticism while preserving the relationships the learner needs to build. As capability grows, the environment can remove predetermined elements and ask the learner to design more of the artifact's own constraints.

Inspect, compare, rebuild

An artifact becomes educational through what happens after it is produced. Inspection asks the learner to identify the assumptions, relations, and choices encoded in it. Comparison places it beside a consequence, rival artifact, canonical explanation, or earlier version. Rebuilding requires the learner to use the identified discrepancy to generate a new construction. Stopping at feedback leaves the most important causal step—the change in the learner's model—unobserved.

Comparison should preserve provenance. If the learner sees an ideal answer and then edits their work until it resembles the ideal, the final artifact mixes recognition, copying, and reconstruction. Versioned artifacts make the transition visible: what was predicted before observation, what criticism was received, which relation changed, and what the learner can later regenerate. A trajectory is epistemically richer than a polished endpoint.

Rebuilding also prevents error correction from becoming purely verbal. Learners often agree with feedback because the corrected sentence feels clear. Asking them to produce a fresh prediction, derive the result by another route, or apply the revised explanation to a counterexample tests whether the criticism has reorganized generation. The relevant question is not whether the learner accepts the correction but whether a changed model now produces different consequences.

Fast loops, not instant answers

Construction benefits from short feedback loops because a learner can still recover the assumptions that produced the artifact. Programming environments, simulations, physical materials, and responsive tutors can make consequences immediate and precise. Rapid iteration increases the number of conjecture–criticism cycles available in a fixed period and allows a learner to explore distinctions that would be too expensive if every attempt required external evaluation.

Speed becomes counterproductive when feedback arrives before commitment or performs the reconstruction automatically. Auto-correction that prevents a malformed step can preserve flow but remove evidence about the learner's syntax model. An AI that rewrites an argument at the first weakness shortens the production cycle while eliminating the explanatory cycle. The valuable latency is long enough for the model to generate something determinate and short enough for the learner to connect the consequence to its cause.

The ideal tool accelerates reality, not the answer key. It makes behavior observable, records versions, enables counterfactual changes, and returns control to the learner. Over time, the artifact sequence should show broader generation under fewer constraints: predictions become more discriminating, explanations integrate more relations, criticism becomes self-initiated, and rebuilding requires less external diagnosis.

What crosses between minds

A teacher begins with a rich internal compression and produces something transmissible: words, equations, diagrams, examples, demonstrations, or tasks. Call that representation X. The learner receives X, not the structure from which X was generated.

X enters a different latent space containing different memories, concepts, routes, and contradictions. The learner must construct their own compression. This is why a universally clear explanation can still fail a particular learner and why the same representation can produce different understandings.

Personalization means construction-aware support

Personalization is not merely changing the reading level or generating more examples. It means choosing the next representation or constraint in response to the model the learner has made visible.

One-to-one tutoring matters when it can observe construction, add the smallest useful constraint, and test what remains after that constraint is removed. If it only delivers increasingly tailored explanations, it can personalize the scaffold while leaving learning untouched.

A design rule for instruction

Judge every explanation by what it asks the learner to generate next. A representation succeeds when it makes a productive construction possible, connects to structures the learner can actually use, and leaves enough open space for the learner to perform the integration.

The question after explaining is not 'Was that clear?' It is 'What can you now build, predict, or explain that makes the new relationship visible?'

Pedagogy and mathetics

Pedagogy traditionally examines the teacher's art: how to sequence, represent, explain, question, motivate, and assess. Mathetics examines the learner's art: how to choose a problem, interpret a representation, generate a conjecture, connect it to prior knowledge, construct an artifact, seek criticism, and reconstruct. The two are complementary, but educational systems have disproportionately formalized the first because teaching actions are visible and administratively controllable while learning actions occur inside a model.

This imbalance changes design. If the unit of analysis is a lesson delivered, improvement means clearer slides, tighter explanations, better examples, and more responsive pacing. If the unit is a construction made, the same lesson remains incomplete until the learner has generated a consequence that could not be read directly from it. The teacher's sequence is evaluated by the learner's subsequent activity, including the productive errors and connections it makes possible.

Mathetics is not a demand that learners teach themselves without inherited knowledge. It is the study and cultivation of what learners must eventually do for themselves even in the presence of excellent teaching. An explanation can enter the process, but the learner must still route it, test it, integrate it, and regenerate its structure later. Education becomes durable as these mathetic operations move from an externally managed routine to the learner's own repertoire.

There is no inverse decompression

An expert explanation is a compression produced from an expert model. Its vocabulary, examples, and omissions rely on relations that feel obvious to the expert because those relations already organize their thought. There is no general inverse function by which a learner can take the compressed representation and reconstruct the same model. Too much information has been omitted, and the learner's prior knowledge supplies different defaults at every ambiguous point.

This is not merely the familiar problem of using jargon. Even perfectly defined words sit inside different networks. 'Force,' 'energy,' 'proof,' 'market,' or 'democracy' can activate school procedures, everyday intuitions, emotional associations, or partially contradictory examples. A representation $X$ is therefore not educational content in isolation; its effect is $\operatorname{Construction}(X, L_B)$, where $L_B$ is this learner's current model. Change the learner and the same $X$ generates a different construction.

The impossibility of exact decompression is productive. Education does not require cloned minds. A learner may build a different internal route that preserves the explanatory relationships needed for prediction and criticism. Multiple representations matter because they create several opportunities for those relationships to be constructed. The criterion is not fidelity to the teacher's phenomenology but whether the learner's model now generates powerful consequences and survives contact with the same constraints.

One-to-one as construction-aware tutoring

The advantage of one-to-one tutoring is often described as personalization, but personalization can mean little more than endless explanatory variation. A construction-aware tutor personalizes the next epistemic demand. It observes the learner's prediction, asks why a relation was selected, notices whether an error is local or structural, and chooses a constraint that leaves the learner responsible for the next meaningful move.

Such a tutor works like a diagnostic instrument. If a learner can continue after a diagram, representation was the bottleneck. If recalling a prior case unlocks the task, routing or connection was missing. If only a modeled solution enables imitation, the generative structure remains largely external. The tutor records not just what the learner got wrong but the minimum intervention under which construction resumed, then later removes that intervention to test whether it became internal.

AI makes high-resolution tutoring economically plausible while making the distinction more urgent. A system can infer confusion and generate custom materials at extraordinary speed, but its default incentive is often to answer helpfully. Mathetic design asks it to optimize for decreasing dependence: elicit before explaining, expose consequences before correcting, make its constraints visible, and increase the learner's share of generation across attempts.

The learner inherits the environment

At first, the environment supplies the loop. It selects a tractable problem, requires a prediction, holds the artifact stable, introduces a critic, and schedules a later reconstruction. With repetition, the learner can appropriate each function. They begin to notice their own comprehension illusion, seek a counterexample, choose a representation, separate a missing fact from a weak explanation, and return to an idea after its context has faded.

This is learning to learn in a precise rather than motivational sense. It is the acquisition of constraints and procedures that improve the learner's own conjecture–criticism cycle. The learner knows how to make uncertainty legible: write the prediction, build the smallest test, compare versions, ask which assumption controls the result, and decide when an external explanation would narrow rather than erase the search.

A mature educational relationship therefore contains its own dissolution. The teacher or system remains valuable as a source of new knowledge, criticism, and collaboration, but no longer needs to manage every act of construction. External constraints fade where internal ones have developed. The learner becomes capable of designing an educational environment around themselves and of recognizing when a supposedly helpful environment is merely producing successful outputs on their behalf.

Local models can coexist

A learner may apply Newtonian mechanics when a classroom problem announces the topic and revert to an Aristotelian intuition in everyday motion. Both structures survive because they are called by different contexts and are rarely forced to determine the same consequence.

This is not simply forgetting. The formal knowledge may be available and independently usable inside one route while remaining disconnected from another. More repetition of the classroom procedure can strengthen the local route without resolving the contradiction.

Relationships are knowledge

Knowing three principles independently does not mean knowing what follows when all three constrain the same problem. The relationships that emerge between them are themselves new knowledge.

Connections are therefore not mnemonic decoration. They are what turn isolated information into a model that can explain and generate. A learner has not simply collected more nodes; they have changed the structure through which future situations are interpreted.

Force models to interact

Use artifacts that require several familiar ideas to jointly determine one output. Remove labels that announce the chapter. Change the setting while preserving the causal structure. Ask the learner to explain apparent contradictions between their own predictions.

The aim is not arbitrary transfer for its own sake. It is to create occasions where locally coherent models can no longer avoid one another, allowing criticism and reconstruction to operate on the relationships between them.

Context routes the model

A context is not decorative background. Its words, objects, goals, emotional stakes, recent examples, and social setting determine which structures are likely to participate in inference. The phrase 'find the force' may activate a formal classroom procedure, while an everyday question about why a bicycle slows may activate an intuition that motion naturally expires. Both answers can be fluent because each is generated inside a route that rarely encounters the other.

The mixture-of-experts analogy captures this selective routing. A system may contain several locally competent structures and a context-sensitive mechanism that activates some more strongly than others. The analogy should not be read as a neurological architecture. Its value is explanatory: adding correct knowledge to one local route does not guarantee that the route will be selected elsewhere, nor that contradictory structures have been reconciled.

Context dependence is not inherently a defect. Experts also use simplified models locally: Newtonian mechanics remains useful where relativistic corrections are negligible, and different social frames foreground different legitimate concerns. The problem is unrecognized dependence. A learner needs to know the boundary of a local model, the cues that activate it, and the conditions under which another explanation should replace or constrain it.

Knowing the parts is not knowing the relation

Curricula often decompose knowledge into separately teachable components and assume coherence will appear when all components have been covered. But the conjunction A + B can generate implications present in neither A nor B alone. Understanding supply and demand separately is not yet understanding how an external constraint changes equilibrium. Knowing acceleration and reference frames separately is not yet knowing which measurements remain invariant. The relation must itself be constructed and criticized.

This is why a learner can pass isolated tests and fail a problem that combines familiar ideas. Nothing has necessarily been forgotten. The assessment is asking for a new model whose components must jointly determine an output. From the learner's perspective, that is new knowledge. Treating the failure as lack of practice with each component can strengthen the nodes while leaving the missing edge untouched.

Dense knowledge is therefore not a warehouse of independent propositions. It is a network in which concepts constrain one another, compress recurring structures, and make distant consequences reachable. Connections are often described as memory aids because they improve retrieval, but their deeper function is generative. They change what follows from what the learner already knows.

Transfer failure as a routing problem

Transfer is frequently described as applying learned knowledge to a new context, but this hides two distinct demands. The learner must possess a relevant structure and must recognize that the present situation calls for it. Instruction can succeed at the first while leaving the second entirely controlled by chapter labels, familiar notation, or teacher prompts. When those cues disappear, the knowledge becomes inert despite being retrievable on request.

Near transfer preserves many routing cues and tests whether a procedure survives modest variation. Farther transfer removes surface similarity or combines the knowledge with another domain. Neither is universally superior; the distance should correspond to the claimed abstraction. If a principle was taught as a general causal relation, evidence should include cases whose vocabulary does not announce it. Failure then reveals either that the abstraction was never constructed or that its routing remains too context-bound.

Designing for transfer means varying contexts during construction, not merely demanding transfer afterward. Learners should compare cases that look similar but differ causally and cases that look different but share a structure. They should name the feature that justifies the analogy, state its limits, and generate a case where the relation would fail. This builds routes around explanatory invariants rather than surface resemblance.

Coherence grows through conflict

A coherent model is not produced by announcing that all knowledge connects. It grows when two locally successful structures are required to answer the same question and cannot both survive unchanged. The friction model and inertia model can coexist until a sequence of cases varies friction while preserving motion. An intuitive statistical story and a formal probability model can coexist until both make explicit predictions about the same sample. Conflict gives the missing relationship a consequence.

Generic advice often fails for the same reason. 'Be confident,' 'work harder,' or 'think critically' names an outcome without representing the context-sensitive models that generate behavior. Useful knowledge connects action to conditions: confident relative to which uncertainty, effort directed at which bottleneck, criticism using which standard? Decontextualized slogans are easy to remember precisely because they compress away the relations needed for intelligent use.

The goal is not a single perfectly consistent worldview. Human action requires local simplification, and some tensions remain unresolved at the frontier of knowledge. The goal is a model capable of noticing when local explanations overlap, identifying the assumptions that permit each one, and constructing an artifact through which their conflict can be criticized. Coherence is an active capacity for integration, not the static absence of contradiction.

Beyond repetition

Repeated exposure can strengthen familiarity while preserving the comprehension illusion. A learner sees the same definition and recognizes it faster. That is useful, but it is not the same as generating an explanation, recognizing relevance, or rebuilding a derivation.

Reconstruction asks the model to do work. The gaps, distortions, and shortcuts in the regenerated artifact reveal what has become stable and what still depends on external support.

Change the constraints

Do not reconstruct the identical output in the identical setting forever. Change the prompt, representation, surface context, available tools, or concepts that must interact. Preserve the underlying knowledge while varying the routes by which it becomes relevant.

This makes spaced reconstruction a test of coherence as well as memory. The learner is not only storing an answer but building a structure that can survive movement.

The cycle

Generate from the current model. Compare with consequence or a stronger explanation. Identify the specific gap. Rebuild. Wait until the supplied context fades. Generate again under a changed constraint.

The cycle is small, but its objective is large: external knowledge becomes cheaper to regenerate internally until the learner can use it as a default part of future thought.

Regeneration, not re-exposure

Re-exposure presents the structure again and asks the learner's model mainly to recognize it. Regeneration removes enough of the structure that the learner must produce it. Both can strengthen memory, but they offer different evidence. Familiarity with a definition, highlighted passage, or completed derivation can rise on every viewing while the learner remains unable to explain the relation from a blank page or identify when it matters.

Generation is not valuable because struggle is intrinsically virtuous. It is valuable because the output depends more heavily on the learner's present organization. Omissions reveal missing relations; distortions reveal how prior models assimilated the idea; false starts reveal the live search space. When the canonical answer remains visible, those signals are overwritten by recognition. The environment becomes less able to distinguish what the learner produced from what the representation supplied.

A reconstruction need not be verbatim. Exact recall is appropriate for symbols, names, or text whose form matters. Explanatory knowledge should often be regenerated through a new artifact: derive rather than recite, draw rather than reread, teach a case rather than quote the rule, or predict an unfamiliar consequence. Variation makes the learner reconstruct the compression rather than only its original surface form.

Let the temporary context fade

Immediately after an explanation, the learner can rely on its wording, order, examples, and activation of relevant concepts. That residual context behaves like a scaffold. A successful reconstruction minutes later is valuable but still underdetermined: the learner may be replaying a recent sequence rather than regenerating the relation from a durable model. Spacing weakens this temporary route and increases the share of the artifact that must be produced by longer-lived structure.

The ideal interval is not simply the longest tolerable delay. If the model can generate nothing, the attempt provides little selective information and may require the entire scaffold to be restored. Productive spacing creates partial reconstruction: enough remains for the learner to form a conjecture, while enough has faded for dependencies and gaps to become visible. The interval should adapt to the knowledge, the learner, and the richness of the artifact.

Forgetting is therefore not merely an enemy to suppress. It can remove accidental support and reveal what has actually become generative. The educational response to a failed reconstruction is not always to repeat the original information. It may be to compare the failed artifact with the earlier one, locate which relation disappeared, and rebuild that relation under a representation that connects more deeply to existing knowledge.

Reconstruct under varied constraints

Repeating the identical prompt can build a highly efficient local route. Varied reconstruction changes the cues while preserving the underlying explanatory demand. The learner might encounter different surface objects, representations, problem goals, tool availability, or combinations of concepts. They must decide which features are invariant and whether the old model still applies. This tests routing and coherence, not only storage.

Variation should be principled. Random novelty can increase difficulty without targeting the model. Useful contrasts include cases with similar surfaces but different causes, different surfaces with the same cause, boundary conditions where a rule stops working, and problems where two familiar principles must interact. Each change asks a specific question about what the learner's construction has generalized.

Constraints can also vary by removal. First reconstruct with a diagram, then without labels; first with a list of candidate principles, then with mixed problem types; first with an AI available for criticism, then with the learner required to generate the critique. The trajectory reveals which functions have moved inside the model and which remain rented from the environment.

Track trajectories, not flashcards

Spaced reconstruction produces a sequence of Artifacts rather than a binary remembered-forgotten record. Version one may preserve the conclusion but omit the mechanism. Version two may restore the mechanism while misplacing a boundary condition. A later transfer may use the relation correctly without its original vocabulary. Comparing these versions reveals the evolving compression far better than a streak count or ease rating alone.

The artifact history should record the conditions of production: delay, cues, tools, context, and criticism received. An increasingly polished output is weak evidence if scaffolding also increased. A rougher explanation produced after a long delay in a new domain may demonstrate more independent structure. The objective is to estimate how the model changes across controlled variations, not to reward cosmetic consistency.

Eventually the schedule itself can become learner-directed. The learner recognizes which ideas still depend on fresh context, chooses artifacts likely to expose those dependencies, and returns before the structure becomes unrecoverable. Spaced reconstruction then ceases to be an external memory algorithm and becomes part of Mathetics: a method by which the learner deliberately converts borrowed explanations into autonomous, criticizable knowledge.