Onset-Commitment in Large Language Models: First-Token Competing-Routes as a Readout of How Committed a Route Is — Set by Training Depth, Moved by Pressure

A language model commits to a route at the very first generated token, and how committed it is can be read there — as the first-token competing-routes count (CR), the number of candidate continuations still in contention at the commit point. This paper develops that onset-commitment readout into a measurement instrument with two orthogonal levers on one dial. Training depth sets the commitment: the deeper a behaviour is trained into the weights, the further into “committed” the onset sits, and the less any later instruction can move it. Semantic pressure moves it: framing alone — “take your time” versus “answer immediately” — re-opens or quenches the onset race, with no clock or token budget for the words to refer to. The readout is invisible in whole-response mean CR and sharp at the onset: pressure shifts first-token CR from ≈2.4 (low) to 1.0 (high), p < 0.001 — a committor spike that decays within two tokens. On the same substrate an in-context disposition leaves the onset openable (gap +0.95) while the fine-tuned version of the same disposition quenches it to 0.0 — depth read directly as instruction-resistance. Susceptibility to the pressure lever follows a preliminary capacity gradient consistent with an inverted-U. The two levers cross-validate each other (depth × pressure) and dissociate from a second, downstream level of deliberation that survives fine-tuning and carries reasoning-accuracy gains (+13.3 points on a 70B model, difficulty-gated). The instrument unifies chain-of-thought, prompt folklore, and friction-theoretic pressure under one readout, giving the quenched-versus-annealed picture of route-competition both its temperature (pressure) and its disorder (training depth). Scope. CR is treated as an instrument for the language model's own decision physics — temperature, annealing, the commit point — with no claim to measure human cognition. Substrate. The competing-routes substrate is developed in the series' foundational papers. Data and code. The pressure-ladder stimuli, the position-wise CR readouts, and the evaluation scripts are released with the publication. v2 update (July 2026). The core measurements are unchanged; this version corrects one directional claim, adds a de-confounded probe, and repositions the paper against a fast-moving literature. The accuracy effect of semantic pressure is corrected to high pressure helps on a hard task under matched framing (a difficulty-gated effect), replacing the earlier direction. A de-confounded onset probe is added in which a legal defer option separates deferring commitment from breaking answer format, and the capacity result is stated honestly as a rise to a mid-capacity peak across two model families rather than an established inverted-U. On positioning, the paper now concedes explicitly that a language model's commitment preceding its verbalisation is established in recent work, for which it claims no priority; the contribution is narrowed to the instrument, the two levers that move the onset (training depth and semantic pressure), the capacity-relative susceptibility of the pressure lever, and the onset-versus-body level separation. The sequential-sampling vocabulary of decision-making is named as the right one for the phenomenon, with the pressure lever framed as a manipulation of decision urgency — not a speed–accuracy manipulation, since accuracy conditional on committing is flat at the onset — and the paper declines to fit an accumulator model while acknowledging the collapsing-bound and urgency-gating accounts of deadline-forced decisions. Builds on v1.