The Physics of Learning: How Race-Architecture Constraints Explain What We Know About Teaching, Communication, and Understanding

Paper 16 · Pødenphant Lund, T. (2026) · Preprint · v5 live on Zenodo

A didactic translation of Friction Theory mechanics to human learning, teaching, and communication. Three traditions of learning science are each empirically excellent and mechanically thin: cognitive load theory and multimedia learning (Sweller, Mayer); desirable difficulties and retrieval practice (Bjork, Roediger, Karpicke); psychological safety and need-prepotency (Maslow, Edmondson). This paper proposes they are local consequences of one substrate-level mechanism: race architecture under bounded compute with irreversible commit-pressure, friction as the cost of unresolved competing routes, hysteresis as encoding-through-loading, and the Net Friction Rule. The claim is substrate-relative, conditioned on the substrate satisfying the race conditions, and is framed as a resource-rational task-analysis in the sense of Lieder and Griffiths (2020).

DOI (concept)	10.5281/zenodo.20416959
DOI (this version)	10.5281/zenodo.22057876 (v5)
Status	v5 live on Zenodo, 2026-08-22
Audience	Education research, instructional design, organisational psychology
Author	Tomas Pødenphant Lund [ORCID]

Three central commitments

Substrate-mechanical account of established didactics

Race-architecture (R1–R3), friction as competing-route cost, hysteresis as encoding-through-loading, and the Net Friction Rule are developed for a non-AI / non-FT-familiar audience. Section 2 builds on race-model concepts already established in cognitive psychology (Ratcliff's drift-diffusion model, Usher and McClelland's leaky competing accumulator, the neural-population race in motor cortex); the framework's addition is to make the three structural conditions explicit and substrate-portable, with commit-pressure as the load-bearing one. From that architecture five classical findings of learning science follow as consequences rather than as separately-validated regularities:

• Working-memory limits — bounded compute on a race architecture implies a small concurrent-route budget for any substrate satisfying R1–R3. The specific human number remains a biological-substrate parameter rather than a universal of race-based processing.
• The testing effect — encoding happens through loading, not from passive exposure. Retrieval is commit-against-alternatives, and the hysteresis it lays down is the encoding.
• Spacing and interleaving effects — commit-against-decay vs commit-against-confusion friction-profiles.
• Expertise reversal — closing an open race (the novice case) and opening a closed one (the expert case) are different substrate-level operations with different costs, which is why the effect-magnitudes are asymmetric: d ≈ +0.5 for low-prior-knowledge learners against d ≈ −0.4 for high-prior-knowledge learners (Tetzlaff et al. 2025). Deriving the asymmetry is the framework's most specific directional contribution.
• Prepotency — safety-field activity competes for race-allocation regardless of substantive-content needs, by the allocation rule rather than by a normative need-hierarchy.

Section 2.8 extends the same architecture to the constructivist precursor beneath all three traditions, offering a candidate physical content for Piaget's operations: a schema is the prior-commitment landscape, assimilation is resolution on an incumbent route, accommodation is a contested race against the committed schema that re-lays the trace, and equilibration is the substrate discharging the friction of an open race. The mapping covers the local step and is scoped away from stage theory and from the long-arc developmental trajectory.

LLM substrates as mechanical mirror

Per-token competing-routes signature, capacity-titration threshold-collapse, volume-asymmetry between race-opening and non-race-opening input, and substrate-graded expertise reversal make the mechanism empirically visible. The competing-routes measure is defined in the paper itself (Section 3.1): the count of candidate next-tokens whose probability exceeds τ = 0.10, averaged over the response, with Paper 1 supplying the full per-token measurement standard and the entropy readout it treats as primary.

Figure 1 is the expertise-reversal mirror, built from Paper 4B's measured data on Llama-3.3-70B, n = 100 per condition. Accuracy runs 73 / 52 / 61 per cent across 0 / 1 / 3 demonstrations while mean CR runs 1.052 / 1.114 / 1.073: competition peaks exactly where accuracy dips, and an independent rerun on a second inference provider agrees to within 0.007 CR. A single demonstration re-opens a strategy-race the zero-shot substrate had already closed. The pattern is substrate-graded: flat at 1.5B (below the capacity at which a strategy-race can run), a classic scaffold at 7B, the U-shaped dip at 70B.

The language-model evidence is positioned as mechanical mirror, not load-bearing. The load-bearing empirical anchor for the human claims is the educational-psychology literature reviewed in Section 1. Paper 4 and Paper 4B supply the mirror measurements; Paper 13 supplies the capacity-titration ladder. Section 3.5 states what does and does not transfer: the placement of mechanism-events transfers, the amplitude does not, and bindings-relief chunking is bottleneck-relative rather than a race-mechanic, an architecture × task × capacity dissociation observed directly on the substrate where architecture can be varied. Paper 4D's reading that trained-in routes resist re-opening under low pressure is held as a candidate barrier interpretation awaiting test, not as a settled result.

Three classroom failure modes from one mechanism

Three distinct classroom failure modes are derived from one substrate-mechanical constraint:

• Dump — raw quantity of race-opening input exceeds the concurrent-race budget. The intervention is volume reduction of race-opening content specifically, and chunking pays off only where a sequential-memory constraint is actually engaged.
• Dilute — correct information at insufficient race-opening density. The intervention is not volume increase but density increase: framing against the learner's defaults, prediction before telling, choice-points that must be committed to.
• Ambiguity-without-commit — multiple strategies sustained throughout without commit-pressure release. The intervention is clarity, not volume in either direction.

The diagnostic-before-intervention rule is what the taxonomy buys: the dump-mode recipe of "say less, more slowly" helps the dump-mode learner and worsens the other two. Section 5.2 sharpens why volume is the wrong variable. On the same substrate and queries, 1,600 random-nonsense filler items cost about 13 percentage points of accuracy, while 9 task-relevant elaboration sentences, about a hundredth of the token volume, cost about 20 in a separate run. The runs start from different baselines and the 20-point drop is not significant at 30 queries, so this hints that race-opening density rather than volume is what counts; it does not show it.

The matched-friction principle (Paper 6) specifies the don't-over-explain corollary as upper-boundary matched-friction violation. Diagnostic protocols and intervention recipes are mode-specific.

Methodological commitments

Programmatic hypothesis

The substrate-level account is presented as a programmatic hypothesis, not a closed theory. Section 2.7 positions Friction Theory relative to predictive-processing and active inference (Friston, Clark, Hohwy), ACT-R (Anderson), and Bayesian cognitive science (Chater, Oaksford); commit-pressure under bounded race-architecture is identified as the load-bearing structural assumption that distinguishes the substrate-level account from these alternatives. Three test cases carry the positioning: the redundancy effect, where the accounts converge in direction and diverge on the positional signature; the testing effect, where the friction account alone predicts that strongly-committed failed retrievals lay hysteresis on the wrong route; and the expertise-reversal asymmetry, which predictive processing does not naturally derive.

Implications

Section 8 includes:

• Curiosity gaps as engineered matched-friction on race-opening (sustain-pressure mechanism), with the testable prediction that optimal gap-size scales with the audience's prior capacity and domain knowledge.
• Forward-framing as pre-commitment placement — a frame supplied before the receiver's route forms incurs neither the exit cost of leaving a sunk investment nor psychological reactance proper; the same content after commitment can incur both.
• Meaning-field mis-diagnosis as dominant organisational learning failure.
• Formal-education abstraction-mismatch as structural feature of context-independent curriculum: the system-level meaning answer does not close the learner's near-horizon meaning-race, and the account predicts a steady baseline of engagement failures rather than episodic ones.
• Dunning-Kruger as a substrate-general default — confident-before-competent falls out of the architecture, modulated by a learned meta-confidence component rather than a special bias. Scoped against the statistical and socio-motivational contributions to the empirical pattern.

Four falsification conditions

1. Scope-conditions violated. A learning substrate that does not satisfy R1–R3. The paper marks this as a scope-specification rather than a sharp disconfirmation.
2. No position-specific friction signature under adequate imaging. Within-subject learning episode, n ≥ 60, pupillometric, EEG or fMRI signal time-locked to commit-event positions, d ≥ 0.3 at p < .01 corrected and replicated. Failure withdraws the positional-friction prediction of Section 9.1.
3. Pure repetition matches varied exposure. Four-condition between-subjects design, n ≥ 80 per cell, delayed recall and transfer at 7 and 28 days. Pure-repetition gain within 0.1 d of varied-exposure gain at 28 days withdraws the variation-versus-repetition prediction and retracts the Section 6.4 claim.
4. A symmetric matched-friction inverted-U. Five or more difficulty levels, n ≥ 40 per level, capacity-stratified, retention and transfer at 14 days. Symmetry around the peak withdraws the wrong-route-hysteresis asymmetry claim of Section 7.2.

Conditions 3 and 4 are the sharp ones. The pattern-level commitment is explicit: the framework survives single nulls, and is in difficulty as a whole if both fail under well-powered pre-registered tests on independent samples.

Connections to other papers in the series

• Paper 0 (BFT) — the four-fields architecture and the canonical home for the behavioural frame. Paper 16's safety-field prepotency and its reactance treatment come from BFT's field-allocation account.
• Paper 1 (A Per-Token Measurement Standard for Route-Competition in Language Models) — the instrument. Paper 16's CR readings rest on this standard and cite it at their point of use; the race conditions R1–R3 and the resource-rational positioning are stated there.
• Paper 4B (Substrates Encode Experience) — encoding-through-loading, and the source of Paper 16's Figure 1 and of the volume-asymmetry and threshold-collapse measurements.
• Paper 4 (Same Content, Wider Track) — the empirical-calibration battery, the second half of the mechanical mirror.
• Paper 6 core (Matched Friction) — the matched-friction-under-hysteresis schema. Paper 16's "don't-over-explain corollary" is the upper-boundary violation.
• Paper 13 (Operational Friction Theory) — race-opening, recursive resolution, and the capacity-titration ladder Section 5.3 draws on.
• Paper 4D (Onset-Commitment) — commit-pressure as an annealing temperature, the substrate reading behind the "give the learner time" corollary of Section 7.3.
• Paper 21 (Mount Stupid in the Machine) — the confidence trajectory Section 8.6 reads as a substrate-general default with a learned modulator on top.

Read the paper

The full paper is on Zenodo (concept DOI 10.5281/zenodo.20416959):

Read on Zenodo → · Plain English version · Dansk version