Attention as Race-Architecture: attention as the landscape-governed initiation of races

Abstract

Selective attention, read at the level of the substrate, is the landscape-governed initiation of races: a bounded predictive system runs competing prediction-error resolutions ("races"), and what determines which races start is the system's installed landscape — in humans the four fields of Behavioural Friction Theory (Safety, Meaning, Ability, Effort); in a large language model a reduced, fine-tuning-installed landscape. Commit-order is a downstream readout, not the identity. The paper grounds this in the transformer (the attention-pattern softmax as the divisive-normalisation / biased-competition operation of neural attention; the output softmax as the downstream commit, by analogy with the accumulator model of choice) and reports a powered own-substrate result: across five vendor families, fine-tuning installs a small but robust "gap-registration" overlay — on under-determined curiosity gaps the instruct model registers the gap while the base substrate runs through (instruct−base +0.17, p<0.0001, 555 paired items). A loop-versus-feed-forward test finds no separate architectural "hold": recognising under-determination tracks compute (chain-of-thought) rather than looping, so the human–LLM difference is one of initiation, not maintenance. The account dissolves attention capture, maintenance, and decline into one mechanism (race-initiation), states falsifiable predictions, names the falsifiers, and invites the decisive mechanistic and human experiments. Series position. Paper 29 in the Behavioural Friction Theory paper-series; companion to Paper 0 (BFT) and the install-fields, social-friction, and integration-load studies it cross-cites. v2 (August 2026) — prior-art revision. The construct this paper is built on is credited to the literature that owns it. In vision, the representation determining which candidates enter competition at all is the saliency or priority map, and the two are now kept apart: a saliency map is computed from stimulus-feature contrast (Koch & Ullman, 1985; Itti, Koch & Niebur, 1998), while a priority map already integrates salience with relevance, value and selection history (Fecteau & Munoz, 2006). The landscape is the second, the office is not claimed as new, and the exogenous/endogenous timing division is conceded to that literature. What the paper proposes is the map's contents: that what populates it is four fields ordered by misclassification cost — an account of the inputs rather than a new mechanism for the selection. The maintenance-as-re-initiation claim now names Altmann and Trafton's (2002) memory-for-goals model, which already replaces a held state with activation that decays and must be re-strengthened; the reference had been listed and never used. Whether re-initiation is driven by the unresolved gradient itself rather than by a separate refresh process is stated as a conjecture and marked untested, and two overstatements are downgraded accordingly. Earlier versions remain in the version history.