Substrates Encode Experience, Not Information: An Encoding-through-Loading Framework with Cross-Substrate Tests in Language Models

Paper 4B in the Friction Theory paper-series — version 3. An eight-experiment empirical paper on language-model substrates that proposes and tests an encoding-through-loading framework, situating it against Cognitive Load Theory (Sweller) and within the recent Multiple Moderated Mediations (Triple-M) framework for prior knowledge × learning (Simonsmeier inputs that open a race the substrate must resolve leave a trace structured by the resolution. The triggers of such races — reactance, pseudo-complexity attribution, capacity-overload, expectation-violation, personal relevance — are sourced from human cognitive psychology; this paper tests their LLM-side analogues and presents evidence consistent with cross-substrate operation. The persistence layer differs: human substrates accumulate cross-session encoding-traces; language models without persistent memory architectures pay friction within a generation but do not propagate it. The mechanism reduces, via predictive-coding, to unreducible prediction-error: random noise habituates to non-signal and costs nothing regardless of volume; ambiguous-but-plausible content cannot be categorised and costs continuously regardless of token-count. Eight predictions tested on multiple LLM substrates: (1) substrate-graded U-curve in expertise-reversal across model size; (2) per-token CR peaks at 1-shot strategy-crossover; (3) elaborated demonstrations reduce friction relative to minimal demonstrations; (4) format-violation produces a 22pp accuracy collapse with measurable CR rise; (5) within-session repetition produces no encoding gain; (6) random-nonsense fillers cost only 13pp across a 1600× volume increase; (7) sweet-spot capacity-graded peaks collapse to overload-floor under pressure; (8) exploratory placement-signature re-analysis identifies trigger-specific patterns. Theoretical positioning. The race-account is positioned within the Triple-M framework as a substrate-level operational specification beneath four of its sixteen mediators (interference, strategy change, skill automatization, transfer). §7.4 elaborates the framework's didactic implications and connects the substrate-mechanism to existing traditions in pragmatic-uptake (Grice, Sperber changes-of-mind per Resulaj/Kiani/Wolpert 2009; motor variability per Wolpert social-correlation caveat per Lorenz/Rauhut/Schweitzer 2011 PNAS. Reactance boundary condition reframed as attenuation via persuasion-knowledge (Friestad & Wright 1994), not sign-reversal. Triple-M section renumbered from §7.4e to §7.4f. Bibliography expanded from 84 to 97 entries: 13 new cites bridging social-learning, persuasion-knowledge, motor-control, and decision-kinematics literatures. The v2 §7.4d receiver-side mirror-friction (communication as pragmatic uptake on top of channel-fidelity) is unchanged in v3. The v1 empirical contribution (eight experiments on Qwen2.5/Llama-3.3/Qwen3-235B/DeepSeek-V3 with the chemistry composition task Y = N × X) is also unchanged.