Operational Friction Theory: Race-Opening, Recursive Resolution, Manifested Behaviour, and Thermodynamic Termination

Abstract

One proposed mechanism, four components, many phenomena. This paper specifies the race-architecture for learning organisms — race-opening, recursive resolution, behaviour-as-a-manifested-route, and thermodynamic termination — and shows it recovers Yerkes-Dodson, the peak-end rule, substrate-graded expertise reversal, compulsive behaviour, intrusion-without-resolution trauma, and task-switching cost as consequences of one proposed mechanism rather than as separate phenomena. A central claim: the operative threshold is race-opening, not commit — commit is the terminus of resolution (or the architectural cycle), carrying no separate tunable decision-bound. Abstract. Friction Theory (Pødenphant Lund 2026b, Paper 1) treats friction as the number of viable continuation routes available to a substrate at each moment of processing. The race-architecture (Pødenphant Lund 2026e, Paper 10) proposes extending this across quantum, classical, biological, and computational regimes. This paper specifies the operational mechanism at four components required for any substrate running the race-architecture: Race-opening condition governed by free-energy gradient relative to current race-state - race-opening is the flow that follows when the gradient is present, not a substrate-decision. Recursive friction-resolution at multiple abstraction-scales simultaneously, with each scale converging at its own commit-deadline and projecting temporally onto the response-token sequence. Behaviour as a manifested resolution-route, taken when friction cannot be discharged through internal routes; restructures the classical input -> cognition -> behaviour model as the special case in which internal routes happen to be available. Sustain-pressure as accumulated blocked-race-pressure observed at three time-scales (within-response, within-session, cross-session), with thermodynamic termination in which non-winning races stop because compute-budget shifts away rather than because they are explicitly rejected. The four components are unified by a single physical metaphor: water flows when there is a gradient, settles in the basin, leaves through whatever channel architecture admits, and stops when there is no gradient left. Placement of mechanism-events is structurally comparable across substrates while amplitude requires per-substrate calibration. Explicit falsification criteria are provided per component, with allowed-vs-forbidden redescriptions making the falsification boundary an explicit contract. Companion papers in the Friction Theory series: Paper 0 (BFT master): 10.5281/zenodo.19462499 (concept) Paper 1 (Friction Theory substrate): 10.5281/zenodo.20012654 (concept) Paper 2 (Capacity Scaling of Encoding-Through-Loading: Application vs. Cloze Asymmetry): 10.5281/zenodo.20013491 (concept) Paper 3 (Friction-Guided Inference: Calibrating Correction Strategies and Abstention from Logprob Signals): 10.5281/zenodo.20014121 Paper 4 (Same Content, Wider Track): 10.5281/zenodo.20059859 (concept) Paper 7 (Forward-modelling in bounded race substrates: agency, self-modelling, and theory of mind): 10.5281/zenodo.20449154 (concept) Paper 10 (Race-architecture): 10.5281/zenodo.20014567 Paper 2F (Compete, Don't Erase: Contradictory Fine-Tuning Masks Rather Than Deletes Knowledge): 10.5281/zenodo.20570433 (concept) Paper 23 (Reading "Friction" in LLM Logprobs: A Substrate-Level Correlate of Present Route-Competition in Social-Moral Judgments): 10.5281/zenodo.20678294 (concept) Paper 28 (Dread without a dreader: anxiety as a non-terminating race, and the collision at the edge of constructive representation): 10.5281/zenodo.20690746 (concept) Paper 29 (Attention as Race-Architecture): 10.5281/zenodo.20703510 (concept) Paper 4B, Paper 6, Paper 8 - in preparation v8 (August 2026) — measurement revision, with one result withdrawn and one exclusion withdrawn. The paper's empirical weight rests on one instrument, the per-token competing-routes count read from a served top-k list, and it now states what that instrument can carry before using it. A new section sets five constraints: levels are run-relative and only contrasts travel, with portability marked where it has been demonstrated and provisional otherwise; the count is truncated at the top of the list and has no resolution on short windows; a thresholded event count carries a response-length term and censors at the generation cap; instrument corrections are per-run rather than per-corpus; and a quoted level inherits every constraint a measured one carries. Serving one model the same prompts through three providers returned mean counts of 1.24, 1.62 and 2.39, which is why the first constraint is first. Re-derived against that standard, the friction-event inverted-U reported in earlier versions on two substrates survives on one and is withdrawn on the other. On the smaller substrate the interior peak survives a threshold sweep at three cut-points, a threshold-free entropy surrogate, a density normalisation and a fixed-window count, and it is reported with those checks rather than as a shape. On the larger one it does not: a counting error fired on every measurement and the curve is confounded with response length, so it is withdrawn, and what the withdrawal costs the framework is stated rather than absorbed. A second-substrate exclusion is withdrawn in the other direction: a near-zero event count previously read as an explained absence is an instrument floor, and reading it as evidence was circular. The accuracy collapse that carries the two-regime structure involves no route-count measurement at all and is unaffected. Three further changes. Concurrency read from where features fall along an output sequence now carries its debt explicitly: Townsend (1972) and Townsend and Nozawa (1995) are cited at the claim, and the placement inference is stated as constrained rather than as settled. Attractor networks (Wang, 2008) are named as the source of the settling account the mechanism's second component uses, with the difference stated — their bifurcation runs along evidence and this one along load. And the single-pass, temperature-zero provenance of the accuracy figures is disclosed beside them, with the published instability rate quoted to size an unmeasured quantity rather than claimed as this dataset's. Earlier versions remain in the version history.