Learning as Anti-Entropic Distinction Building

Module 4 is the framework's most developed application domain. Its purpose is not to survey learning science in distinction-vocabulary — most of its sections inherit content already established in cognitive psychology — but to give a theoretical account of why a specific deployed language-learning methodology works. That methodology is SSi (Say Something in Welsh, Spanish, and other languages), deployed since 2009 across tens of thousands of learners. Its design rules are developed under the name HISE (High-Intensity Speaking Exercises). The module's central theoretical move is §4.2, which derives an action functional from the two axioms and identifies HISE's design rules as action-reducing choices along the trajectory to conversational production.

The module's structure reflects this. §4.1 introduces learning as distinction refinement. §4.2 develops the variational spine. §4.3 returns to the cognitive primitive — the same/different judgement — on which the variational structure runs. §4.4–4.9 interpret individual phenomena through the distinction vocabulary that the variational structure makes specific. §4.10 addresses empirical consistency; §4.11 addresses educational practice; §4.12 closes with cross-module connections.

The Phenomenon to Be Explained

Seventeen years of SSi deployment produce a specific learner phenomenology the theoretical account has to explain:

• Production-threshold crossing. Learners generate spontaneous speech rather than merely recognize or translate. The shift from recognition to production is precisely what most language-learning apps fail to produce.
• Imperfection tolerance. Learners will produce output while still imperfect. Perfectionism does not calcify into avoidance behaviour.
• Generative-overlap error patterns. Errors at the end of acquisition are confusions between neighbouring chunks (similar-sounding lexical items, near-synonymous M-LEGOs), not structural inability to produce.
• Retention across long gaps. Acquired production capacity persists without continuous practice, beyond what flashcard-based methods deliver.
• Typological robustness. The methodology holds across languages as varied as Welsh, Spanish, Japanese, and Irish — phonologically, grammatically, and typologically different target languages.
• Low explicit metalinguistic knowledge relative to production capacity. Learners often cannot articulate the grammar rules they reliably follow in spontaneous speech.

These patterns are the explanandum. The framework has to account for them as consequences of the two axioms; if it cannot, the theoretical account is inadequate. They are not testimonials. They are the data the module has to fit.

The Theoretical Account

Recall the two axioms developed in Module 0:

Axiom 1: All distinctions accessible to OLUs cost energy. IMPORTS Landauer; scoped to OLU-accessibility — see §0.3

Axiom 2: All observers-like-us (OLUs) operate under finite energy budgets.

§4.2 develops the core theoretical move: the two axioms combine to give any learning trajectory a natural action functional $S=\int E[n,\dot{n}]dt$ with units of energy × time. The least-action trajectory is the one that reaches conversational production with minimum total resource expenditure. HISE's design rules — fixed LEGO form, M-LEGOs with bundled particles, BUILD/USE duality, production-before-recognition, hierarchical tiling, same/different as organizing primitive — can each be read as local reductions of the cost integrand. The functional follows from the axioms; the fit of the empirically-evolved rules to it is interpretive. The account is therefore quantitative in form, not merely descriptive.

The explanandum patterns are interpretable as consequences. Production-threshold crossing can be read as the effect of BUILD/USE duality forcing automatization during novelty. Generative-overlap errors are what sprint-length trajectories produce when they cover the production region of $\mathcal{N}$ without dwelling off-trajectory. Retention tracks automatization migrating distinctions into low-maintenance circuits. Typological robustness follows from the axioms not depending on features of any specific language. These are observed outcomes the account fits, not theorems forced by it; the interpretive account is worked through in §4.2 and elaborated in the phenomenon-specific sections that follow.

The Anti-Entropic Direction

The variational structure specifies the trajectory. The anti-entropic framing specifies its thermodynamic direction. Boundaries — distinctions held in a learner's network — decay without continuous energy investment. Learning is the active maintenance and extension of distinction patterns against that decay.

This framing is Module 4's specific instance of the broader thermodynamic framework developed in Module 7. The action functional of §4.2 is the learning-domain cost functional; the anti-entropic framing is the direction the trajectory runs against.

Epistemic Status

Claims in this module distribute across three tags:

• DERIVED — The action functional $S=\int Edt$ and its consequences. These follow from Axioms 1 and 2 without additional assumptions.
• INTERPRETED — The identification of specific learning phenomena (chunking, automatization, spacing, transfer, skill acquisition) with action-reducing mechanisms within the trajectory. The interpretations are consistent with established cognitive psychology but the framework does not independently derive the phenomena. The identification of HISE with an approximate least-action pedagogy is similarly interpreted — empirically supported by long deployment, not formally proven.
• IMPORTED — Landauer's principle grounds Axiom 1. Neuroscience observations about metabolic costs of cognition (the ~20W brain-power figure, the 30–50% metabolic reduction under practice) support the energetic claims but are not derived.

The module's contribution is not that it derives cognitive science from scratch. Its contribution is a variational structure that gives a quantitative, falsifiable theoretical account of a specific deployed methodology, with predictions about short-timescale acquisition trajectories testable against transcript data.

Module 4 Section Map

The module's argument depends on §4.2. The surrounding sections elaborate and interpret, but the theoretical weight rests on the action functional derived there. A reader short on time can read §4.0 and §4.2 and have the core claim in hand; the rest is elaboration and cross-reference.

Section 4.0 announced the module's explanandum and named §4.2 as its central theoretical move. This section introduces learning's local structure — the kinds of modifications a learner makes to their distinction network over time. The four modes developed below (acquiring, sharpening, reorganizing, automating) will each be recognized in §4.2 as local moves within the trajectory $n(t)$ through configuration space $\mathcal{N}$ — increments of $\dot{n}$ contributing to the cost integrand $E[n,\dot{n}]$.

Through the distinction lens, learning is one thing: an OLU refining its own capacity to tell things apart. That refinement is anti-entropic — it draws boundaries and holds them, against the pull toward dissolution. It takes several forms:

1. Acquiring new distinctions: Learning to differentiate states that were previously undistinguished (the wine novice learning to distinguish Bordeaux from Burgundy)
2. Sharpening existing distinctions: Increasing the reliability or precision of existing boundaries (the radiologist learning to distinguish subtle tissue abnormalities)
3. Reorganizing distinction hierarchies: Restructuring which distinctions are primary and which are derived (the physics student reconceiving motion in terms of energy rather than force)
4. Automating distinctions: Moving distinction operations from high-energy conscious processing to low-energy automatic circuits (the musician no longer consciously distinguishing individual notes while sight-reading)

All four share one thermodynamic signature: each shifts the relationship between energy spent and distinctions made. So learning is not the accumulation of facts. It is an efficiency gain — more distinctions for the same energy, or the same distinctions held more reliably for less.

The Thermodynamic Definition of Learning

Formally, if $E(d,t)$ is the energy cost of making distinction $d$ at time $t$, and $Q(d,t)$ is the quality (reliability, precision) of that distinction, then learning has occurred between $t_1$ and $t_2$ when:

Or equivalently:

The ratio $Q/E$ represents distinction-making efficiency, and learning is the process of increasing this efficiency over time.

The two axioms established in Module 0 — that distinctions cost energy, and that observers-like-us operate on finite energy budgets — combine to give any learning trajectory a natural cost functional. That functional has units of energy × time. In classical mechanics this quantity is called action, and trajectories that extremize action are the ones physical systems actually follow. Here we show that language acquisition admits an analogous variational structure — the action functional itself follows from the two axioms — and that the design rules of High-Intensity Speaking Exercises (HISE) can be interpreted as choices that approximately minimize action along the path from zero competence to conversational production.

The argument in this section is self-contained. It rests only on the two axioms and on a minimal specification of what a distinction network is. Nothing that follows requires physics beyond the Landauer grounding already imported into Axiom 1.

4.2.1 From Two Axioms to a Cost Functional

Let the distinction network of a learner at time $t$ be represented by a state $n(t)$ in a configuration space $\mathcal{N}$. The precise structure of $\mathcal{N}$ is developed in Module 1; for present purposes it suffices that $\mathcal{N}$ is high-dimensional and that each coordinate encodes one distinction. A distinction may be a phonological contrast, a lexical item (the LEGO ga), a morphosyntactic chunk bundling a noun with its particle (日本語を), or a hierarchical M-LEGO whose internal structure has already been acquired. Each coordinate carries values for the robustness of that distinction and for its degree of automatization.

A learner's trajectory through $\mathcal{N}$ is a continuous path $n:[0,T]\to \mathcal{N}$ with initial state $n(0)=n_{\text{zero}}$ (no target-language distinctions) and terminal state $n(T)=n^{\ast }$, where $n^{\ast }$ denotes any network state sufficient to generate and comprehend spontaneous speech in the target language. The velocity $\dot{n}(t)$ describes the rate at which the network is being modified — new distinctions being added, existing distinctions being strengthened, some distinctions migrating from high-energy explicit circuits to low-energy automatic ones, others decaying from disuse.

From Axiom 1, every distinction currently held and every distinction currently being modified incurs an instantaneous energy cost. Write this as a cost rate $E[n(t),\dot{n}(t)]$ with two components:

$E_{\text{maint}}(n)$ is the cost of holding the current network in place against decay — roughly proportional to the number of distinctions being actively maintained, weighted by their individual maintenance costs. $E_{\text{work}}(\dot{n})$ is the cost of modifying the network: adding distinctions, strengthening existing ones, migrating distinctions between circuit types. Both components are non-negative under Axiom 1.

From Axiom 2, the learner cannot sustain arbitrarily high $E[n,\dot{n}]$ for arbitrarily long. Let $B$ denote the learner's sustainable energy-budget rate (dimensions of power). Over duration $T$, total available resource is bounded by $B\cdot T$.

The natural resource functional for a learning trajectory is then the integrated cost rate along the path:

$S$ has units of energy × time. This is the action of the trajectory — the total resource cost of the specific path taken from $n_{\text{zero}}$ to $n(T)$. Axiom 2 constrains feasibility: a trajectory is realizable only if $S\le B\cdot T$. The structure $S=\int Edt$ has been inherited directly from the two axioms. Nothing has been imported from physics that was not already present in the Landauer grounding of Axiom 1.

4.2.2 The Pedagogical Optimization Problem

Given $n_{\text{zero}}$, a target $n^{\ast }$, and a time horizon $T$, a pedagogy is a prescription for the trajectory $n(t)$ through $\mathcal{N}$: which distinctions are introduced when, which are reinforced, which are held explicit, which migrate to automatic circuits. Different pedagogies produce different trajectories. The trajectory that minimizes $S$ subject to reaching $n^{\ast }$ within $T$ is the least-action acquisition trajectory.

This is a variational problem. Its full solution is not tractable analytically — $\mathcal{N}$ is too high-dimensional and the cost functional too coarsely specified for that. But the framework can identify excess action in any candidate pedagogy: regions of the trajectory where $E[n(t),\dot{n}(t)]$ is higher than it needs to be to approach $n^{\ast }$. There are three characteristic sources.

Off-trajectory maintenance

The pedagogy leads the learner through coordinates of $\mathcal{N}$ that are not on the path to $n^{\ast }$. The network accumulates maintenance cost for distinctions that will not be used in the target. Reading-first methods build orthographic distinctions; orthography and articulation occupy largely independent regions of $\mathcal{N}$ for the oral-production target. Energy spent on orthographic maintenance is energy not approaching $n^{\ast }$.

Premature distinction-load

The pedagogy increases $E_{\text{maint}}(n)$ early, requiring the learner to hold many distinctions simultaneously before any of them has stabilized or found a production anchor. A complete verb-conjugation table demands the learner maintain distinctions across tense × person × number × mood for each verb — a combinatorially large set of distinctions with no immediate output use. The network pays maintenance cost on all of them without any of them yet doing work.

High-action automatization paths

The pedagogy leaves distinctions in high-maintenance explicit circuits for longer than necessary before migrating them to low-maintenance automatic ones. Automatization itself has a work cost ($E_{\text{work}}$), but the time-integrated saving in $E_{\text{maint}}$ more than compensates. Methods that delay the explicit-to-automatic transition — recognition-heavy approaches that never force production, for instance — pay extended maintenance on circuits that could have been migrated earlier.

Each of these is a way a trajectory can accumulate $S$ without approaching $n^{\ast }$. A pedagogy that eliminates all three is approximately least-action for the target. A pedagogy that eliminates none is typically an order of magnitude more expensive — which is what "learning a language takes years" often describes in practice.

4.2.3 HISE as Action-Minimizing Pedagogy

The design rules of HISE — refined over seventeen years of deployment across tens of thousands of learners — can each be read as choices that reduce action along the trajectory from $n_{\text{zero}}$ to conversational production. The directional discipline matters: the rules came first, refined empirically, and the framework interprets them after the fact as local reductions of the cost integrand — each of which either shrinks $E_{\text{maint}}(n)$, redirects $E_{\text{work}}(\dot{n})$ toward on-trajectory regions, or both. This is an interpretive fit, not a derivation of the rules from the axioms.

Each rule is a reduction, not a pedagogical feature. The rules are individually derivable from the cost functional and are jointly consistent: none conflicts with any other, and their combined reduction — measured over the deployed methodology — is an order of magnitude or more relative to the alternative designs discussed in §4.2.2.

4.2.4 The 10-Day Sprint as Empirical Probe

A learning trajectory with $T$ measured in years reaches $n^{\ast }$ under many pedagogies, because even high-action trajectories eventually accumulate enough work on the network to approach the target. Long-duration trials are therefore poor probes of trajectory choice: they integrate over so much time that differences in instantaneous $E[n,\dot{n}]$ are smoothed out by total effort.

Short-duration trials discriminate. A 10-day sprint to conversational production has $T$ on the order of 80–100 sustained attention-hours. If two pedagogies reach qualitatively different $n(T)$ over this window — one producing spontaneous speech, the other only producing recognition — the difference is almost entirely attributable to the trajectory: which coordinates of $\mathcal{N}$ have been traversed and how much $E_{\text{work}}$ has been directed at production-relevant circuits rather than elsewhere.

Two such sprints from the deployed methodology are informative here:

• A 10-day Japanese sprint culminating in a 20-minute video call with a native Japanese speaker. Japanese is agglutinative, particle-heavy, phonologically unfamiliar to English speakers, and uses a non-Latin script. The target $n^{\ast }$ is particularly distant from $n_{\text{zero}}$ along multiple orthogonal axes of $\mathcal{N}$.
• A 10-day Irish sprint culminating in a 15-minute radio interview. Irish has VSO word order, initial consonant mutations, and phonology with no close English analog. A radio interview as endpoint is a particularly demanding probe — no pauses, public stakes, spontaneous response.

The variational framing predicts specific features of the production output at $n(T)$:

1. Generative-overlap errors, not paradigm-gap silence. The trajectory has covered the coordinates of $\mathcal{N}$ relevant to spontaneous production. Errors at $n(T)$ reflect local ambiguity between neighboring LEGOs (similar-sounding lexical items, near-synonymous chunks), not structural absence of the capacity to produce.
2. Flexible M-LEGO use. Learned chunks appear in positions and combinations not seen during training. Flexibility is the operational signature of successful chunking — the chunk has become a low-maintenance higher-order distinction rather than a fixed surface pattern.
3. Disfluency located at distinction-selection points, not at automatized junctures. Pauses occur where the network is still choosing between neighboring LEGOs — the correct place for a pause, signalling genuine selection — rather than where the learner is reconstructing a chunk from decay.
4. Low explicit metalinguistic knowledge relative to production capacity. The trajectory has spent $E_{\text{work}}$ on automatization rather than on maintaining an explicit grammar module. A learner asked "what is the grammatical rule for $X$?" may answer poorly even while reliably producing correct $X$ in spontaneous speech.

These are predictions, not descriptions. The transcripts of the Japanese and Irish sprints — analyzed for error typology, chunk flexibility, disfluency location, and the gap between production and explicit knowledge — either exhibit these features or they do not. The framework can be confronted with the data and the prediction can fail. That is the test.

4.2.5 Falsification Conditions

The variational structure itself is derived from the axioms and cannot be falsified independently of them: if Axiom 1 holds and Axiom 2 holds, then $S=\int Edt$ is the natural cost functional of the trajectory, full stop. What is falsifiable is the identification of HISE with an approximate least-action trajectory. Three specific conditions would require revising this identification:

1. A pedagogy is demonstrated that reaches $n^{\ast }$ with measurably lower $S$ than HISE — comparable total attention-hours, equal or better final production, different design rules. The variational structure stands; the specific rules identified as action-reducing do not, and the methodology updates.
2. Sprint transcripts show paradigm-gap errors rather than generative-overlap errors. The prediction that BUILD/USE duality drives early automatization fails. Either the mechanism is mis-described or the design rule is less action-reducing than claimed.
3. Learners show high explicit metalinguistic knowledge and low production capacity. The trajectory has spent $E_{\text{work}}$ off-trajectory. The claim that HISE directs action toward production circuits is incorrect, and the methodology is building something other than what the framework claims it builds.

Conditions 2 and 3 are empirically accessible with the sprint transcripts in hand. Condition 1 requires comparative deployment evidence against alternative methodologies at matched intensity; it is accessible but has not been systematically measured. The audit reporting on Conditions 2 and 3, with citation-level evidence from both transcripts, is in §4.2.5.1 below.

4.2.5.1 Sprint Findings (May 2026)

The two sprints named in §4.2.4 have now been audited against the four predictions and the §4.2.5 falsification conditions. Clean transcripts of both — the Irish radio interview and the Japanese video-call segment — exist. The cross-sprint summary is reported here; within-sprint detail and citation-level evidence are in the accompanying evidence pack (`docs/evidence/sprint-findings-may-2026.md`).

Across both recordings, all four §4.2.4 predictions are positively visible: generative-overlap errors at chunk-selection points rather than paradigm-gap silence; flexible M-LEGO use across novel positions and combinations; disfluency located at selection points rather than mid-chunk; activity-talk dominant over rule-talk when learners describe what they did. The two §4.2.5 falsification conditions assessable from transcripts — F2 (paradigm-gap silence) and F3 (high metalinguistic knowledge / low production capacity) — are absent in both recordings.

The signal beyond within-sprint consistency is the cross-typology convergence. Irish and Japanese diverge on word order (VSO / SOV), morphology (initial mutation / agglutinative suffixes), particle density, phonology, and script. The same error / disfluency / M-LEGO signature appears in both. This is the §4.0 typological-robustness claim in its strongest available form: cross-typological convergence on the same production signature, despite divergent target-language structures, under the same methodology.

The framework was confronted with both named test recordings and the predictions did not fail; the predictions were positively visible across both. The interpreted-status story about HISE survives confrontation with its named test data, across two typologically divergent target languages.

4.2.6 On the Physics Analogy

The action structure here is not metaphor. In classical mechanics, action $S=\int Ldt$ selects the trajectory a physical system actually follows, by Hamilton's variational principle $\delta S=0$. In language acquisition, action $S=\int Edt$ is the resource cost of a learning trajectory — and here the important difference — nothing automatically selects the least-action trajectory. Nature selects via $\delta S=0$ as a law of motion. In learning, the trajectory is chosen by the pedagogy designer, or emerges from the learner's own allocation of attention. Good pedagogies approximate the least-action trajectory either through theoretical insight or through long empirical selection pressure on what methods survive across generations of learners.

What is shared between classical mechanics and language acquisition is the mathematical structure: a cost functional integrated over paths in a configuration space, with a distinguished extremal trajectory. The framework does not import this structure from physics; it derives the cost functional from its own two axioms. That the resulting optimization problem shares mathematical form with the variational principles of classical mechanics is a convergence of cost-functional-over-paths mathematics, not a borrowed theorem.

This convergence is what makes the framework theoretically distinctive in the learning domain. Cognitive science describes phenomena (chunking, automatization, spacing). Pedagogical research catalogs what works (distributed practice, retrieval practice). Neither identifies the underlying cost functional that these phenomena are optimizing. The framework does — as a direct consequence of two axioms about distinction-making under resource constraints. The remaining sections of Module 4 trace each of these phenomena back to its role in minimizing action along the acquisition trajectory.

Section 4.2 derived the action functional $S=\int Edt$ and identified the least-action trajectory as the target any pedagogy should approximate. That structure assumed a ground-level operation — the distinction-making step encoded in each coordinate of $\mathcal{N}$ — without examining what that operation looks like from the learner's side. This section returns to it.

The framework's originating observation was that the single cognitive move underlying all learning is a same/different judgment. When a learner encounters something new, they ask "is this the same as something I already know, or different?" This is the primitive operation on which the variational structure of §4.2 runs — the step whose repeated local increments, under energy constraints, give rise to the trajectories and cost integrands of acquisition.

All learning reduces to processing similarity and difference. This is not just a psychological observation — it follows from the nature of distinction itself. The same/different judgement is not like distinction-making. It is distinction-making, caught in its most immediate, observable form.

Recall that distinction is the primitive operation that enables all cognition. When encountering any new input, an OLU must perform two complementary evaluations:

These are the two sides of boundary-drawing. Together, they locate the new input within the existing distinction network.

Why This Duality Is Fundamental

Consider learning to recognize a new category, such as learning to identify a particular bird species. The learner must process:

• Differences that separate this species from similar species (distinctive markings, size, behavior)
• Similarities that unite members of this species (common features across individuals)
• Differences that separate all birds from non-birds (wings, feathers, flight)
• Similarities that place this in the broader category of birds

At every level, the duality is running. Even the most basic perceptual learning — pulling figure from ground — demands both moves at once: what holds the figure the same across its extent, and what sets it apart from the ground. Same and different are never separate operations. They are two readings of one boundary.

Energy Implications of the Duality

This asymmetry explains several well-documented learning phenomena:

Learning works this asymmetry. It builds distinction structures that spend as little as possible on difference-processing the task does not need, while staying sharp to the differences the task does. The aim is not to notice everything. It is to notice what matters, cheaply.

Chunking, the grouping of multiple elements into single units, emerges directly from thermodynamic constraints on distinction-making. It represents a fundamental anti-entropic strategy: instead of maintaining many separate distinctions (high entropy, high energy), the system organizes them into hierarchical structures (low entropy, lower maintenance cost).

The Energy Cost of Maintaining Distinctions

From Axiom 1, maintaining each distinction costs energy. For an OLU with energy budget $B$ and per-distinction maintenance cost $c$, the maximum number of simultaneously maintainable distinctions is:

For OLUs, working-memory capacity is most economically described as energy-budget-limited ($N\approx B/c$) rather than slot-limited — an interpretation consistent with the chunking data, not a derivation of the specific capacity figure.

How Chunking Reduces Energy Cost

Consider memorizing a phone number: 5-5-5-1-2-3-4. Maintaining seven distinct digit-boundaries costs $7c$. But if chunked as 555-1234, only two boundaries need active maintenance ($2c$), with the internal structure stored as compressed patterns that require minimal energy until accessed.

The energy savings from chunking can be expressed as:

When $E_{\text{internal}}\ll (n-1)\times E_{\text{boundaries}}$, chunking provides significant efficiency gains.

Chunking as Learning

Learning often proceeds by building increasingly sophisticated chunking structures:

1. Initial stage: Many low-level distinctions maintained simultaneously (high energy cost)
2. Intermediate stage: Some distinctions grouped into chunks (reduced energy cost)
3. Expert stage: Complex hierarchical chunking (minimal energy for maximum distinction capacity)

Chess masters exemplify this progression. Where novices see individual pieces requiring separate distinctions, masters see meaningful configurations—entire opening sequences, strategic patterns—each a single chunk encompassing what would require dozens of distinctions for a novice.

In variational terms (§4.2), chunking reduces $E_{\text{maint}}(n)$ along the trajectory: the same production range is sustained with fewer concurrent active distinctions, so the integrand $E[n,\dot{n}]$ is lower at every subsequent time step. The action reduction compounds — each new chunk lowers maintenance cost for all trajectory time that follows its formation, not just at the moment of chunking. HISE's hierarchical M-LEGO tiling is the chunking operation specialized to language acquisition.

As skills are practiced, they migrate from conscious to automatic processing. Within the thermodynamic framework, this represents a transfer of distinction-making from high-energy to low-energy neural circuits. This migration is anti-entropic stabilization: the distinction patterns become more resistant to decay by embedding in specialized, low-maintenance neural structures.

The Neural Energy Hierarchy

Different brain regions have different energy requirements for the same computational operations:

Learning involves progressively shifting distinction-making operations down this hierarchy.

The Automatization Process

When a skill is first learned, prefrontal cortex is heavily engaged, making explicit distinctions about each step. With practice:

1. Repeated patterns are encoded in basal ganglia (habit formation)
2. Motor sequences are stored in cerebellum
3. Sensory-motor mappings become direct in primary cortices

Each migration reduces the energy cost of the distinction-making operation.

Consistency with Known Metabolic Findings

If automatization is energy optimization, then practiced tasks should show measurably reduced metabolic activity in higher brain regions. The framework is consistent with the established finding that they do — a finding documented independently of this framework, so what follows is retrodiction/consistency, not a confirmed forward prediction:

• fMRI studies show reduced prefrontal activation for practiced tasks
• Glucose consumption decreases in task-relevant regions with expertise
• Oxygen uptake becomes more efficient with practice

The energy reduction is not slight. Studies of motor learning report reductions on the order of 30-50% in metabolic cost for practiced vs. novel movements (figures vary by task and measure). This is consistent with thermodynamic optimization.

The Cost of Consciousness

Conscious processing is energetically expensive because it requires maintaining explicit distinctions in working memory, which has high metabolic overhead. Automatization bypasses this by encoding distinctions in specialized circuits that operate without conscious access.

Automatization frees consciousness for new learning while maintaining previously learned distinctions at low energy cost. This explains the characteristic signature of expertise: experts make fewer conscious decisions while achieving superior outcomes.

In variational terms (§4.2), automatization is an $E_{\text{work}}$ investment that buys a permanent reduction in $E_{\text{maint}}$. Migrating a distinction from explicit circuits (high $E_{\text{maint}}$) to automatic ones (low $E_{\text{maint}}$) costs work at the moment of migration but saves maintenance cost for all subsequent trajectory time. The integrated saving more than offsets the one-time work cost, which is why HISE's BUILD/USE duality (§4.2.3) forces automatization to begin during novelty rather than after: the earlier the migration, the larger the integrated saving along the trajectory to $n^{\ast }$.

Forgetting is entropy winning. It is not a system failure but the natural thermodynamic direction reasserting itself when anti-entropic investment ceases. From the dynamism implication of Section 0.3, we know that all boundaries require continuous energy investment to maintain. When energy is withdrawn, boundaries decay—entropy reclaims the ordered patterns that learning had built.

The Mechanism of Forgetting

Each stored distinction—each memory, each learned pattern—is a maintained boundary. Maintaining these boundaries has ongoing metabolic cost. When:

1. Energy is redirected to other boundaries (attention shifts), or
2. Total energy budget decreases (fatigue, sleep), or
3. Maintenance signals weaken (disuse)

The boundary begins to decay. The distinction that was once sharp becomes fuzzy, then undetectable.

Forgetting Rate Depends on Boundary Type

Different distinctions have different maintenance costs and different decay rates:

So we forget a phone number in minutes but never forget how to ride a bicycle. The bicycle lives in low-maintenance cerebellar circuits and costs almost nothing to keep. The phone number leans on continuous hippocampal-cortical upkeep — withdraw it, and the boundary dissolves.

The Wisdom of Forgetting

Under a finite budget, forgetting is not a flaw. It is housekeeping. An OLU that held every distinction it ever made would spend its whole budget on upkeep and have nothing left to learn anything new. Forgetting clears maintenance resources for the next round of building.

This is why sleep involves active forgetting. During sleep, the brain selectively prunes distinctions, maintaining those with high utility while allowing low-utility distinctions to decay. The result is more efficient use of finite boundary-maintenance resources.

The Spacing Effect Explained

The spacing effect—where distributed practice produces better retention than massed practice—emerges from thermodynamic considerations:

1. Massed practice: Maintains distinctions at high intensity briefly, then allows full decay
2. Spaced practice: Allows partial decay, then re-energizes boundaries just before complete decay

Each re-energizing event after partial decay strengthens the boundary by reconstruction. The learner has to re-draw what was going fuzzy, and re-drawing encodes more robustly than mere holding. The effort of recovery is the point — not a cost to be avoided, but the mechanism that does the work.

Formally: If $E_m$ is maintenance energy and $E_r$ is reconstruction energy ($E_r>E_m$), then:

But the spaced approach produces more durable boundaries because reconstruction engages deeper encoding mechanisms than mere maintenance.

In variational terms (§4.2), forgetting is not action-reducing in itself — it is the natural direction $E_{\text{maint}}$ points when energy is withdrawn. But selective forgetting is action-reducing: allowing low-utility distinctions to decay while reinforcing high-utility ones frees $E_{\text{maint}}$ capacity for distinctions on the critical path to $n^{\ast }$. Sleep-pruning is the system offloading maintenance cost from distinctions that do not contribute to the trajectory's forward progress. The spacing effect similarly redistributes $E_{\text{work}}$ across time so that reconstruction events fall where decay has opened the largest saving.

Neural plasticity — the brain reshaping itself in response to experience — is where learning becomes physical. In the thermodynamic framing, plasticity is the mechanism that reorganizes a distinction network for efficiency. It is the physical implementation of anti-entropic pattern building: the brain rebuilding itself to hold more distinctions for less energy. The metaphor and the mechanism are the same thing.

Synaptic Plasticity as Boundary Adjustment

Individual synapses implement distinction boundaries. Synaptic strengthening (long-term potentiation) sharpens a boundary; synaptic weakening (long-term depression) relaxes a boundary.

1. Strengthen boundaries that are repeatedly activated: Frequently made distinctions become sharper and more reliable
2. Weaken boundaries that are rarely activated: Rarely made distinctions fade, freeing energy for other purposes
3. Create new boundaries for novel patterns: New experiences drive the formation of new distinction structures
4. Eliminate boundaries that consume energy without providing useful distinctions: Inefficient structures are pruned away

This is the Hebbian rule — "neurons that fire together wire together" — read through a different lens. Not arbitrary association, but thermodynamic efficiency: boundaries that earn their keep are kept, and the ones that do not are let go.

Structural Plasticity as Network Reorganization

Beyond synaptic changes, the brain exhibits structural plasticity: dendrites grow and retract, axons extend to new targets, entire networks are pruned or expanded. This represents more radical boundary reorganization:

• Dendritic growth: Creating capacity for new distinctions
• Axonal extension: Connecting previously separate distinction networks
• Network pruning: Eliminating inefficient or redundant boundaries
• Myelination: Reducing energy cost of frequently used distinction pathways

Critical Periods and Plasticity Windows

Why is some learning so much easier at certain ages? The thermodynamic framing offers an answer, and it is about cost, not magic:

Early in development, energy allocation to distinction-making is highly flexible. The system has not yet committed to particular boundary structures. This makes radical reorganization possible but also makes the system vulnerable to inefficient encodings.

As development proceeds, the most frequently used boundaries become stabilized in low-energy circuits. This reduces plasticity but improves efficiency for established distinctions.

Once the window closes, the same reorganization means dismantling structures that have already set — and that costs far more. This is why a child takes on a first language without apparent effort, while an adult learning a second one has to pay, in sustained work, for boundaries that were once nearly free.

In variational terms (§4.2), plasticity is the physical substrate on which $\dot{n}$ — the rate at which the network is modified — is implemented. Synaptic changes are the biological mechanism of $E_{\text{work}}$ expenditure; structural changes (dendritic growth, axonal extension, pruning, myelination) are the mechanisms by which the network topology is reshaped to reduce future $E_{\text{maint}}$. The critical-period phenomenon follows directly: during windows when structural plasticity is cheap, large reductions in future $E_{\text{maint}}$ are attainable at low $E_{\text{work}}$ cost. After the window closes, the same reduction costs substantially more work — the trajectory is constrained to a higher-action region of $\mathcal{N}$.

Transfer is what happens when learning one thing makes the next thing cheaper. It works whenever the distinction structures built for the first domain still apply to the second. So transfer is not a separate faculty — it is prior anti-entropic investment paying out again: energy spent building patterns in domain A that domain B can borrow rather than rebuild.

Positive Transfer: Efficient Boundary Reuse

Where domain A and domain B share underlying distinction structure, building it once for A discounts the cost of B. The learner does not draw new boundaries; it reuses the ones already drawn.

• Learning one Romance language facilitates learning others (shared phonological and grammatical distinctions)
• Learning mathematics facilitates learning physics (shared quantitative and relational distinctions)
• Learning one musical instrument facilitates learning others (shared auditory-motor distinctions)

The transfer efficiency depends on the overlap between distinction structures:

When overlap is high, transfer is substantial. When overlap is low, there is little advantage.

Negative Transfer: Boundary Interference

Sometimes existing distinctions interfere with new learning. This occurs when domain A uses distinctions that are incompatible with domain B.

• English speakers learning Mandarin tones (must learn to distinguish what English treats as same)
• Tennis players learning badminton (similar but different motor distinctions)
• Experts in one field approaching another with inappropriate categories

Negative transfer has a clear thermodynamic signature. The learner now has to spend energy suppressing a boundary while drawing its replacement — fighting one distinction to make room for another. That is more expensive than starting from nothing. Prior knowledge is not always a head start; sometimes it is a debt.

Meta-Learning: Learning to Transfer

Experienced learners develop meta-cognitive distinctions about their own distinction-making. They learn to:

1. Recognize when existing boundaries are applicable
2. Inhibit inappropriate boundaries
3. Identify structural similarities across domains
4. Build modular distinction structures that maximize transfer potential

Meta-learning makes every later round of learning cheaper — a higher-order optimization that compounds. Learn to learn well once, and the saving is paid out on everything that follows.

In variational terms (§4.2), transfer is action reduction by reuse. A distinction structure already present in the learner's network supports acquisition in a second domain at reduced $E_{\text{work}}$ cost — the learner does not have to construct from scratch what overlaps with existing structure. The transfer-efficiency ratio defined above is precisely the factor by which integrated $E_{\text{work}}$ in the new domain is reduced. Negative transfer is the opposite: existing distinctions on the wrong trajectory add $E_{\text{work}}$ cost to suppress before the new structure can form.

Skill acquisition runs a characteristic shape, and the thermodynamic framing reads it as progressive anti-entropic optimization. The novice starts with costly, disordered distinction patterns; the expert holds lean, well-organized ones. The road between them is anti-entropic pattern building, made visible.

The Learning Curve

Performance typically improves rapidly at first, then more slowly, approaching but never quite reaching perfect performance. This follows from energy optimization:

1. Initial phase: Many inefficient distinctions; high energy cost; large room for improvement
2. Middle phase: Progressive optimization; chunking and automatization reduce energy cost; improvement continues but at slower rate
3. Asymptotic phase: Most possible optimizations achieved; remaining improvements require increasingly costly reorganization

The familiar learning curve — power law or exponential — falls out of this. Each optimization uses up some of the room left to optimize, so the returns diminish. The plateau is not the learner stalling; it is the trajectory running out of cheap moves.

Plateaus and Breakthroughs

Skill acquisition often shows plateaus—periods of little apparent progress—followed by sudden breakthroughs. The thermodynamic framework explains this:

A plateau is the price of admission to the next structure. Moving from one configuration to another means tearing down working boundaries first — a cost paid up front, before any benefit shows. The breakthrough comes when enough reorganization energy has built up to finish the move. The flat stretch is not nothing happening; it is the bill being paid.

Expertise as Deep Optimization

Experts in a domain have achieved extensive thermodynamic optimization of their distinction structures:

1. Task-relevant distinctions are highly refined: Precise, reliable, and low-energy
2. Task-irrelevant distinctions have been pruned: No wasted maintenance energy
3. Distinction hierarchies are optimally organized: Efficient access and retrieval
4. Automatic circuits handle routine operations: Minimal conscious energy required

This explains the characteristic advantages of expertise:

• Faster processing: Lower energy per distinction means quicker completion
• Higher accuracy: Better-calibrated boundaries yield more reliable distinctions
• Greater capacity: More distinctions maintainable within the same energy budget
• Flexible adaptation: Well-organized hierarchies enable rapid reconfiguration for novel situations

There is nothing mysterious in skill. It is the visible trajectory of an energy-constrained system getting better at telling things apart, over time. Every expert was once a novice whose network had not yet been optimized; every novice is a system carrying enormous unspent potential, waiting only for the investment.

In variational terms (§4.2), skill acquisition is the large-scale trajectory from $n_{\text{zero}}$ to $n^{\ast }$ in a given domain. The power-law or exponential learning curve reflects that early in the trajectory many action-reducing moves remain available; later, the trajectory is already near a local optimum in $\mathcal{N}$ and further reductions require structural reorganization (which itself has $E_{\text{work}}$ cost). Plateaus are regions where the trajectory has saturated a local optimum; breakthroughs are reorganizations that move the trajectory to a different, globally-more-efficient local optimum. Expertise is the state of having executed most of the domain's available action-reducing moves.

The framework makes specific, testable claims about learning. Be clear about what kind they are: not bold new predictions, but consistency demonstrations — showing that the framework lines up with what neuroscience and cognitive psychology already know. Beyond that theoretical fit, it has been borne out in working learning systems deployed in practice.

Confirmed Consistency Points

The following empirical findings are directly consistent with the thermodynamic framework:

1. Practiced tasks require less energy: fMRI and PET studies consistently show reduced glucose metabolism for practiced versus novel tasks. This directly confirms that learning reduces the energy cost of distinction-making.
2. Automatization shifts activity to lower-energy circuits: Motor learning studies show progressive migration from prefrontal to basal ganglia to cerebellar circuits, each with lower metabolic cost.
3. Sleep consolidation involves selective forgetting: Memory research confirms that sleep preferentially maintains useful memories while allowing others to decay, consistent with energy optimization.
4. Chunking increases effective capacity: Working memory studies show that chunked material permits more information retention, as predicted by reduced boundary-maintenance costs.
5. Spacing effects enhance retention: Decades of educational research confirms the spacing effect, explained by the thermodynamics of boundary reconstruction.

Every one of these findings was established independently of the distinction framework, by people not working in it. That they fall out naturally from our two axioms — $E(d)>0$ for all distinctions $d$, and finite energy budgets for all OLUs — shows the framework is coherent with what is already known. It does not prove the framework; it shows the framework does not break against the evidence.

Novel Predictions for Future Investigation

Beyond recovering known results, the framework generates predictions that could be tested with current neuroscience methods:

The Metabolic Signature of Learning

The energy reduction from learning is not subtle. Motor-learning studies report 30-50% drops in metabolic cost for practiced versus novel movements. That is thermodynamic optimization you can measure — the same distinction-making, done for substantially less energy.

A reduction of that size says something. Learning is not about capacity, it is about efficiency. The expert's edge is rarely that they can do what the novice cannot — it is that they do what the novice does, for a fraction of the metabolic cost.

Practical Implementation: Working Learning Systems

A note on the direction of inference. The deployed pedagogies referenced throughout this module — SSi (since 2009), Zenjin, Alexander — were built and refined on operational design rules well before the theoretical articulation here. They are the explanandum the framework attempts to explain (per §4.0 and §4.2), not engineered implementations of principles articulated after the fact:

• SSi (Say Something in Welsh, Spanish, and other languages) — The spacing-effect-driven scheduling and BUILD/USE production-first design predate the variational account. The framework offers an interpretation of why such scheduling sits near a least-action trajectory; SSi is not derived from the framework.
• Zenjin — The hierarchical chunking architecture was designed on operational pedagogical reasoning. The framework offers vocabulary in which to read its design rules as $E_{\text{maint}}$-reductions; the design rules themselves are prior.
• Alexander — The adaptive distinction-building approach reflects long-running pedagogical practice. The framework provides interpretive vocabulary, not a derivation of the system's design.

These deployments are evidence the explanandum exists and has the shape §4.0 describes. They are not testimonials, and the framework is not validated by their existence — at most, the consistency between the variational account and the design rules that work in practice is suggestive. The directional discipline is important: pedagogy → theoretical interpretation, not the reverse.

What Would Falsify the Framework

Honesty demands we name what would count against the thermodynamic account, not just what fits it:

• Finding that practiced tasks consistently require more energy than novel tasks would directly contradict the framework.
• Evidence that learning occurs without any metabolic correlates would challenge the energy-cost basis.
• Discovering that forgetting rates are entirely independent of maintenance energy costs would undermine our account of memory decay.
• If chunking provided no working memory advantages despite reduced boundary count, our mechanism would be falsified.

To date none of that contradicting evidence has appeared. The fit between the framework and the findings does not make it true — nothing earns that — but it does establish it as a viable candidate for unifying learning science under thermodynamic principles.

The framework has consequences for how we teach. If learning is anti-entropic pattern building — drawing stable distinction structures and holding them against decay — then good teaching is whatever makes that process easier. The principles below are consistent with the design rules of long-deployed systems (SSi since 2009, Zenjin, Alexander), whose practice §4.10 treats as the explanandum the framework sets out to explain. Keep the direction straight: pedagogy first, theoretical reading after — never the reverse.

Design for Chunking

Present material in chunks that can be consolidated before adding more. The structure of presentation should facilitate boundary-grouping, not arbitrary sequences.

• Organize content into meaningful units that correspond to natural distinction boundaries
• Allow consolidation time between chunks---rushing prevents boundary integration
• Make hierarchical structure explicit so learners can build nested chunk architectures
• Use consistent patterns within chunks to reduce internal distinction costs

Support Automatization

Provide sufficient practice to move core distinctions to automatic circuits before building more complex skills on top of them. Premature advancement requires maintaining too many high-energy explicit distinctions simultaneously.

• Ensure foundational skills are automatized before introducing dependent skills
• Monitor signs of cognitive overload---they indicate too many high-energy distinctions
• Use varied practice contexts to promote robust automatization
• Allow sufficient repetition for neural migration to lower-energy circuits

Leverage Spacing

Distribute practice to take advantage of the reconstruction effect. Interleave topics to prevent massed practice on any single distinction set.

The spacing effect emerges from boundary dynamics: partial decay followed by reconstruction creates more durable encodings than continuous maintenance. Formally:

Though spaced practice requires more total energy, the reconstruction events ($E_r$) engage deeper encoding mechanisms, producing boundaries that are more resistant to decay.

Manage Energy Load

Distinction-making burns metabolic resources — this is not a figure of speech. A session that runs the budget dry will fail, however well designed. Rest, nutrition, and pacing are not soft extras; they set the ceiling on how much can be learned at all.

• Schedule demanding learning during periods of high metabolic availability
• Include rest intervals to allow partial recovery of energy resources
• Recognize that glucose depletion impairs high-energy distinction operations
• Balance cognitive demands across a learning session to avoid exhaustion

Build Transfer Bridges

Explicitly connect new material to existing distinction structures. The more overlap the learner can recognize, the lower the energy cost of new learning.

• Activate relevant prior knowledge before introducing new content
• Make structural analogies explicit---show learners how new distinctions map onto familiar ones
• Teach general principles that transfer across multiple domains
• Address negative transfer by explicitly marking where old distinctions do not apply

Teach Meta-Learning

Help learners develop distinctions about their own distinction-making. Meta-cognitive awareness enables strategic optimization of learning processes.

• Teach learners to recognize their own cognitive load and adjust accordingly
• Develop awareness of when existing distinctions are helping versus hindering
• Build skills for identifying structural similarities across domains
• Foster understanding of how practice, spacing, and rest affect learning outcomes

Summary: The Thermodynamic Pedagogy

On this view, teaching is not the maximizing of information transfer. It is the optimizing of how efficiently distinctions get built and held. The teacher's job is to shape experience so the learner can grow a well-organized, low-maintenance network — properly chunked, properly automatized. Not to pour in more, but to make what goes in cost less to keep.

This module's central theoretical move is §4.2: the two axioms combine to give any learning trajectory a natural action functional $S=\int Edt$ (this functional follows from the axioms), and the empirically-evolved design rules of HISE can be interpreted as choices that approximately minimize action along the path to conversational production. The surrounding sections situate the move. §4.3 names the cognitive primitive the variational structure runs on (same/different as the ground-level operation). §4.6 names its thermodynamic direction (anti-entropic boundary maintenance). §4.4, §4.5, §4.7–4.9 interpret individual phenomena — chunking, automatization, plasticity, transfer, skill acquisition — through the distinction vocabulary that the variational structure makes specific.

The seventeen-year deployment of SSi, and the subsequent systems built on the same principles (Zenjin, Alexander), is the empirical phenomenon the theoretical account has to explain. The 10-day Japanese and Irish sprints analyzed in §4.2.4 are the short-timescale probes where trajectory choice dominates over total effort. The framework's predictive claims about sprint output are stated there; its falsification conditions are stated in §4.2.5.

This section traces how the module connects to the rest of the treatise.

Connections to Foundation (Module 0)

The two axioms of Module 0 — distinctions cost energy, observers operate under finite budgets — are the full input to Module 4's theoretical content. No additional assumptions are required. Axiom 1 gives the cost rate $E[n,\dot{n}]$; Axiom 2 bounds the feasible region of trajectories; the action functional and its least-action trajectory follow as direct consequences. The transcendental grounding of distinction-making in Module 0 extends to the learning domain: the same operation that underlies thought underlies acquisition.

Connections to Formalization (Module 1)

Module 1 formalizes the distinction operator and the OLU as a system maintaining distinctions under energy constraints. Module 4 specializes the framework to learning: the configuration space $\mathcal{N}$ of §4.2 is a learning-specific instance of the distinction-network structure developed in Module 1. The velocity $\dot{n}(t)$ is the rate at which the OLU modifies its own network — the subject of Module 1's dynamism discussion, now given a cost-functional context.

Connections to Mathematics (Module 2)

Mathematical learning, in the framework's vocabulary, is the acquisition of notation that compresses distinction structure into a form maintainable at low energy cost. The effectiveness of mathematical formalism lies in its chunking efficiency: a single equation $E=m{c}^2$ encapsulates distinction structure that would otherwise require extended verbal maintenance. Mathematical expertise is efficient distinction compression.

Connections to Consciousness (Module 3)

Conscious learning is the explicit maintenance and manipulation of distinction boundaries under attentional allocation — in Module 3's vocabulary, self-referential distinction-making directed at the learner's own network. The relationship is bidirectional: consciousness enables deliberate boundary manipulation; acquired distinctions structure the content of experience; automatization migrates practiced distinctions below the conscious threshold, freeing attentional resources for new acquisition.

Connections to Quantum Mechanics (Module 5)

The uncertainty relations of quantum mechanics and the energy–precision trade-offs of learning share structure: both are consequences of finite-energy observers attempting to maintain distinctions with bounded resource. Module 5's interpretation of quantum formalism through resource-constrained observation runs parallel to Module 4's interpretation of acquisition through resource-constrained network modification.

Connections to Thermodynamics (Module 7)

Module 7 interprets thermodynamics through the distinction framework: entropy as distinction-decay, the Second Law as the natural direction of boundary dispersion without energy input, temperature and free energy as indices of distinction-maintenance capacity. Module 4's anti-entropic framing is the specific manifestation of Module 7's principles in the learning domain. The learner is a far-from-equilibrium system sustaining distinction patterns against decay through continuous energy investment. The action functional $S=\int Edt$ derived in §4.2 is the learning-domain cost functional that Module 7's broader thermodynamic framework subsumes.

The Unifying Theme

Learning is not a capacity added to physical systems. It is what energy-constrained distinction-making looks like when the system modifies its own network to reduce future action along its trajectory. Every learning phenomenon — from the infant's first word to the master's refined expertise — is a manifestation of the same process: boundary-drawing under resource constraints, trajectories through configuration space, cost functionals being minimized locally under an attention economy.

This is why learning follows predictable patterns, has consistent neural substrates, and can be enhanced by understanding its principles. The variational structure is not an analogy to physics — it is the direct consequence of the two axioms applied to a network-modifying observer. Physics and learning share the structure because both are instances of finite-resource systems traversing a configuration space under a cost functional. That is the unifying theme of the treatise, made concrete in the domain where its predictions can be most directly tested.

This section carries the detail of the canonical statement of Least-Time Learning. The front page of that statement — the principle, the trinity, and each of the parts below in summary — is `docs/canonical/least-time-learning.md`. That node is the single place the idea is stated; Configuration Economics, Zenjin and SSi reference it rather than restating it, and this section is where its working is set out. Where Module 4 already carries the material, this section cross-references rather than duplicates: the action functional and the least-action reading of HISE in §4.2, same/different as the ground-level operation in §4.3, deployed pedagogy as explanandum in §4.0 and §4.10, educational practice in §4.11.

The Principle

Learning is optimised by minimising the learner's total effort-time over a distinction network. Every distinction the learner is asked to hold must be doing work at the moment it is charged — never before, and never alongside a second one smuggled in unpaid.

The name is a ruling, not a flourish. The everyday word carries the whole sense; the technical cousin charges a parse toll. The theory's own name passes the theory's own pricing rule.

least time in some ways is more apt than least action — it's easier to understand… time is not smuggling in any confusions, unlike action as a term with a distinct meaning in physics.

— Tom Cassidy, 2026-08-24

The Trinity, as It Bears on Learning

The trinity is the frame this project hangs on, and it is why Least-Time Learning is a node of Distinction as Primitive rather than a standalone theory of teaching.

• Ontology: distinction networks. What there is, for an observer like us, is distinctions and their relations. In the learning domain this is the central move: the subject is a network, not a syllabus, and the learner is a network being modified. §4.2's configuration space $\mathcal{N}$ is that claim made formal; §4.3 names the ground-level operation it runs on.
• Epistemology: persistence of stable distinction patterns. What counts as knowing is which patterns hold up over time under a bounded budget — not access to absolutes. In the learning domain this is what makes forgetting (§4.6) and automatization (§4.5) the same subject as acquisition: knowing is a maintained pattern, and maintenance has a rate.
• Ethics: selection over configurations, denominated against heat death. What to do is a matter of which configurations get selected, against the one denominator that does not move. In the learning domain this is what makes a curriculum an ethical object and not merely a convenience: a curriculum is a selection over configurations of a person's network, and spending a learner's seconds badly is a real cost, not a stylistic one.

Stating the trinity at the head of the canonical node closes half the de-duplication problem in one stroke, and leaves the Script — the machinery of one's own choosing — as the only essential idea still homeless, which sharpens that question rather than burying it.

The Ontology It Stands On

Subjects are distinction networks (the Distinction Project claim). A fact is an agreement — a first-party experience collapsed into a shared label. A concept is a compression — a derivable relation that regenerates families of distinctions. This is why the principle is one thing across languages, maths and the sciences rather than three separately discovered teaching tricks: SSi drills the agreement layer, Zenjin drills the structural layer, and both are minimising the same integral through different strata of the one network.

The Quantity and Its Three Cost Terms

The minimised quantity is learner effort-time: cognitive work integrated over the seconds it takes. This is a real functional, not a metaphor, because all three costs a curriculum can charge are already denominated in it.

• The charge. Acquiring a distinction costs effort over the seconds of acquisition, and the cost grows with distinction distance from what the learner already owns. This is why order matters at all.
• The idle charge. Holding an unused distinction costs a maintenance rate integrated over the interval until it does work. This term is literally effort × time — and it is why the doing-work rule is a derivation, not an axiom: "doing work at the moment charged" is simply the condition that drives the idle interval to zero.
• The debt. A false bridge charges the acquisition cost plus a forced-effort term downstream: every future use of the corrupted region costs suppression work, and eventually unlearning. Debt is worse than idle charge because idle charge integrates until first use and stops, while debt integrates until repaid. The canonical specimen: teaching the alphabet song before decoding — letter names actively interfere with reading, so the child pays for the learning and then pays again to un-learn it. No existing theory prices this.

Every term is measurable. Time-at-the-blink is directly observed; hesitation, wrong-direction misses and "I don't know" rates are the effort signal. The weights are set by telemetry, not by the armchair — "we only get there by testing."

The Shape: Fermat, Not Hamilton

The correct physics ancestor is Fermat's principle: light takes the path of least time through a medium of varying refractive index. The subject is the medium; distinction density is the refractive index; the learner is the ray; a well-built curriculum is the geodesic. "The geodesic to the most intangible domain" — said of the electricity path before this statement existed — was the theory speaking early.

The Two Laws

• When (doing work at introduction): a distinction is charged at the moment it does work, because any earlier accrues idle charge and any later blocks the path.
• What next (order by assembly distance): the next item is one distinction from the covered graph — question N is question N-1's composition plus exactly one new distinction. Categories and topics are packaging; the binding order is local to each node's ancestors ("it's a bloody graph — you can go in any direction you like"). Paths are examples; the graph is the artefact.

And the whole method in one line, the objective function already ruled canon-grade: minimum admitted, maximum minted — introduce and state the fewest things, then mint the most from them. Least time is what that heuristic serves.

The Corollary Instruments

The instruments already built are corollaries, not separate rules. Each was built before the principle was stated, and each falls out of it.

• One distinction per blink — keeps each measurement's cost term pure.
• Parse cost is non-domain cost — a slow parse is effort-time the question has no right to charge, and it contaminates the telemetry that sets the weights.
• The comparison declaration — names the two states each node distinguishes, so the ask is derived, not authored.
• The pricing rule for names and notation — vocabulary as cache policy: buy a word or a symbol only when re-derivation cost × frequency beats storage. "Oscillate" was bought; "mass" and "velocity" never arrive; .5 is minted from halving £7 before it is christened. Notation is a purchase, and mistimed purchases are where the parse toll hides (notation-is-the-toll: confirmed by the maths audit).
• Calibration at the first discriminating rung — placement itself is a least-time move: no token spent on questions that do not discriminate.

The Evidence

• SSi — the THAT. Fifteen years of never introducing a form before the sentence that needs it, working and continuously improved, with the underlying principle unstated. This statement states it. (The deployment and its phenomenology are set out in §4.0 and §4.10.)
• Phonics — the derived-not-empirical result: decoding-first falls out of the functional (letter-name-first is a debt purchase), and the empirical literature then agrees.
• The physics grammar passing two subjects' audits: the same six rules (objects first, one distinction per step, names last, felt experience, law-level analogy, pricing) verified on physics and then found already obeyed in the maths graph's bones.
• The measurement: 494 / 770 / 846 claims (physics / chemistry / biology) — the first time the felt "size" of a subject has a mechanism: fact-count measures how much of a subject is agreement layer versus mintable from compressions. Internal control, which carries the weight honestly: genetics reads physics-shaped within biology, holding the boards' enumeration style constant.

Honest Ancestry: Four Neighbours, the Comparison Cutting Both Ways

What is genuinely new, tightly: (1) the variational form — SSi's method, phonics and Zenjin's ordering derived from one principle; (2) false bridges as debt — mislearning priced for the first time; (3) names and notation as purchases under cache policy; (4) fact-count as a measurement of a subject's generativity.

The Predictions: What Makes It a Theory Rather Than a Philosophy

• The discriminating experiment. Same items, graph-ordered versus school-ordered. CLT predicts no difference (intrinsic load fixed); least-time predicts a large one. The outcome is a number: total learner-seconds to the same owned frontier. Cheap to run on Zenjin's own telemetry, and it doubles as the first measurement of the cost weights.
• Error localisation. Wrong answers cluster at nodes more than one distinction from the covered graph. A wrong answer at distance one indicts the item, not the learner — the clarity-of-ask ruling is this prediction already in use.
• Debt is measurable. Letter-names-first children show decodable interference years later, and unlearning costs more than never-having-learned.
• Notation timing. Operation-before-symbol cohorts beat symbol-first cohorts on transfer specifically, not recall — the hoLo prediction.
• Conversion collapse. Learners taught fractions, decimals and percentages as one comparison find conversion nearly free; three-topics learners find it expensive. Directly testable on Evan.
• The generativity gradient. Biology's 326 single-board rows — pure agreement layer, no assembly discount — are the slowest-learned content in the estate; genetics the fastest corner of biology.

Stated Limits and Open Edges

The cost weights are empirical; telemetry sets them. Per the standing epistemics: "even if we make a ruling, it's likely to be a heuristic — we're making this up as we go along." Held firmly, revised cheaply.

This frame covers the wisdom leg of contribution (wisdom × willingness × wherewithal). Willingness — the machinery of one's own choosing — is a learnable domain with its own prospective graph (the Script), deliberately outside this statement.

Where This Node Lives

This is a Distinction Project node — the pedagogy chapter of the distinction thesis. Its front page is `docs/canonical/least-time-learning.md`; this section is its detail. Configuration Economics, Zenjin and SSi reference the node and do not restate it, citing it as Least-Time Learning — the canonical statement, `docs/canonical/least-time-learning.md`.