Abstract. Behavioural Friction Theory (BFT; Pødenphant Lund 2026a) describes four functional fields — safety, meaning, ability, effort — through which human cognition computes decision costs. The theory is biologically grounded. It has not previously been connected to fundamental physical principles, and it has not been tested outside biological substrates. This paper does both, and in doing so proposes Friction Theory (FT): a candidate substrate-universal framework of which BFT would be the biological instantiation. The central move is to reframe friction itself. Friction is not a psychological state or a metaphor from mechanics. It is the information-processing cost of probabilistic computation, formally connected to thermodynamic free energy minimisation through Ortega Ratcliff 1978; Usher Cisek non-biological race substrates are predicted to exhibit friction without fields. The formal relationship is BFT ⊂ FT (§8). The framework is tested on three large language model architectures (Cogito-671B, Qwen3-235B, Llama-3.3-70B) with a paired Qwen2.5-32B base-vs-instruct comparison. Seven signatures of in-session budget allocation are recovered: iterative-pipeline commit-timing that falls near the secretary problem's 1/e ≈ 36.8% optimum; parse-vs-generate phase decomposition; constructive versus destructive friction types; friction profiles as cognitive fingerprints; mode-shift entry and exit costs; reactance that scales with RLHF intensity across models — a cross-model association for which a controlled within-model test is specified but not yet run; and trailing-task forgetting under high mid-task load (d = 1.2, the strongest cross-model effect). Cross-substrate data from Saigusa et al. (2008) and Laibson (1997) place these findings in a proposed six-substrate temporal-horizon gradient: LLMs (inference-bounded), Physarum (hours), C. elegans (minutes to 40 hours), Drosophila (seconds to 24 hours), cephalopods (minutes to days), and mammalian brains (seconds to decades). Several implications follow. BFT's four fields are reframed as evolutionary derivatives of the safety field under mortality, mobility, and metabolic constraint. Classical cognitive biases — anchoring, confirmation bias, sunk cost, loss aversion — are reinterpreted as thermodynamic necessities in any race architecture. A deeper unification developed in §5.8 proposes that Kahneman's peak-end memory bias, dopaminergic reward-prediction-error signalling, Friston's free energy minimisation, and attention-weighted saliency in transformer LLMs (ρ = +0.17 between token surprise and downstream attention, substantially position-mediated) are substrate-specific signatures of a single mechanism: surprise-weighted state retention. Hysteresis is the structural precondition for learning in any bounded probabilistic system. The framework's empirical support to date is in language-model substrates, where friction is directly measurable as competing token routes in logprob distributions. Its extension to biological substrates is presented as a falsifiable research programme — a set of substrate-specific predictions for slime mould, nematodes, insects, cephalopods, and mammalian brains — not as an established result. This paper lays out the architecture, the language-model evidence assembled to date, and the predictions that invite cross-disciplinary test and falsification. Paper 1 in the Friction Theory paper-series. The foundational substrate theory: friction as the information-processing cost of probabilistic computation in any race architecture. Behavioural Friction Theory (Paper 0) is recovered as the biological instantiation. Version 4 (2026-05-30). Empirical correction and companion-cite update: §5.8 / §7 P19 (attention-saliency prediction): downgraded from CONFIRMED to PARTIALLY CONFIRMED. A confound-controlled re-analysis finds the +0.17 rank correlation between per-token surprise and downstream attention-saliency is substantially position-mediated: after partialling out normalised position and token length, ρ drops to −0.068 (p = 0.047); z(CR) is non-significant in multiple regression (p = 0.49) while z(position) dominates (+0.580, p < 0.001). Cross-architecture replication with explicit position controls is required for robust confirmation. Paper 14 cliff-event (Pødenphant Lund 2026q, DOI 10.5281/zenodo.20217712) added as companion-cite and load-bearing LLM-substrate evidence for race-architecture — cross-architecture replication at p < 1e−17 on FT-substrates (Qwen2.5-7B + Mistral-7B-v0.3), not subject to the position confound. §9 research-programme bullet updated: P19 replication target now specified as position-controlled cross-architecture test. Version 3 (2026-05-24). Major epistemic reframe — from "established substrate-universal theory" to "proposed candidate substrate-universal research programme": Title subtitle changed to Toward a Substrate-Universal Theory of Bounded Computation. Abstract, §1, §3, §5.2, §5.7, §9.0, §9.1 reframed: the framework is proposed and invites cross-disciplinary test; seven LLM signatures are the evidence assembled to date; biological-substrate extension is explicitly a research programme, not an established result. Companion-citation demotions: Paper 4B cross-cites (§5.6.5, §6.1b, §9.5) reframed from load-bearing empirical anchors to see-also companion references; Paper 6 in-text cites demoted to companion-analysis framing; Paper 10 in-prep cites replaced with direct claim framing. §A2 polish: discovery-narrative cleanups (§A2-i, ii, iii) and minor wording improvements (§A2-v, vi) throughout. Bibliography currency: live companion DOIs refreshed throughout. Version 2 (2026-05-20). §7b Kubo (1966) fluctuation–dissipation anchor; three Paper 4B empirical cross-cites (§5.6.5, §6.1b, §9.5); §9.5 open question on the discounting measure for memoryless substrates; bibliography updated; editorial revision throughout. Companion papers in the Friction Theory series: Paper 0 (Behavioural Friction Theory master): 10.5281/zenodo.19462499 Paper 2 (Capacity scaling): 10.5281/zenodo.20013491 Paper 2B (In-context learning as working memory, fine-tuning as long-term memory): 10.5281/zenodo.20145218 Paper 3 (Friction-guided inference): 10.5281/zenodo.20014121 Paper 5 (Field-theoretic taxonomy of emotions): 10.5281/zenodo.20058825 Paper 10 (Race architecture across substrates): 10.5281/zenodo.20014567 Paper 13 (Operational Friction Theory): 10.5281/zenodo.20059876 Paper 14 (Logic as Reactance): 10.5281/zenodo.20217712 Paper 4 / 4B / 6 / 8 / 8b / 8c / 9 — in preparation
Tomas Pødenphant Lund (Sat,) studied this question.