The Race Must Go On: Encoding-Frames Reposition Route-Competition Onset in LLM Generation

Abstract

A short purpose-frame placed before a fact-block produces a significant first-token route-competition effect in a non-RLHF base model that is null in the matched RLHF-instruct model; the most parsimonious reading is that instruction-tuning compresses the competing-route distribution, leaving the frame's onset signal readable on the base and quenched on the instruct. Friction theory reads each generation step as a race between candidate token-routes (competing-route / CR competition in the output distribution); this paper measures where a short encoding-frame repositions that competition. The asymmetry replicates across three model families (Llama-3.1-8B, Mistral-7B-v0.3, Gemma-2-9b; accuracy-matched). The effect is not RLHF-created — it is present before instruction-tuning and absent after. That reading is scoped to the onset race the paper measures: it is not a claim about the model's total generative diversity, which is not measured here and need not move in the same direction. The frame repositions onset competition rather than reducing it by a fixed amount: the direction is task-dependent (it lowers onset CR on a chain task and raises it on a maximally-specified cloze task), and a frame placed after the data cannot un-fire a race that has already opened. Around that core the paper characterizes the same race-positioning mechanism across frame-types, tasks, and substrates (peak-shift, mismatched-frame reactance, combined-frame interference, length-as-race-cost, and a capacity-gated chunking dissociation). Part of the Friction Theory research programme; companion papers are linked in this record's related identifiers. v2 update (July 2026). The empirical results are unchanged. This version situates the paper against the prior literature on prompt framing and sets a boundary on how far its central reading may be taken. A new related-work discussion documents that the phenomenon — that a prompt's framing systematically moves a language model's behaviour — is independently established at the output level in earlier work, for which no priority is claimed; the paper's contribution is distinguished as the measurement of where in the generation the frame acts and on which models the effect stays readable. The reading that instruction-tuning compresses the competing-route distribution is explicitly scoped to the onset race the paper measures and is not extended to a claim about total generative diversity, a distinction drawn against a recent analysis of how fine-tuning restructures uncertainty. Reciprocal cross-references to companion papers are added, and companion citations updated to their current versions. Builds on v1. v3 (August 2026) — reporting-precision and self-containment revision. The instruction-tuning reading is stated as an observed pattern, with the frame-by-tuning interaction marked explicitly as untested; serving and provenance details are given at the tables that depend on them; a lawfulness formulation is downgraded to a hypothesis; abstract-level summaries are reconciled with the per-cell results they describe; companion material is condensed to pointers. Editorial pass. Earlier versions remain in the version history.