Two Candidate Readouts of a Proposed Common Race: Effort-Value Attribution and Commit-Position as Substrate Signatures of Race-Architecture

Engineering notes

Engineering notes will be added by the aipentium editorial team.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为大语言模型、生成式AI、ChatGPT相关技术、计算机视觉、深度学习等高价值论文补充中文说明。

Original abstract

Six effort-value biases — the IKEA effect, the endowment effect, the sunk-cost fallacy, the generation effect, effort justification, and the effort heuristic — and the widening of a language model’s commit-position under instruction-tuning are proposed here as two readouts of one race-architecture, both measurable from token logprobs alone. It is a research direction with explicit falsifiers, not a settled result. A race-architecture under bounded resources — parallel evaluation of candidates, accumulation under resource limits, and an irreversible commit — leaves a substrate state during and after each resolution. The paper proposes two candidate substrate signatures of that state and specifies the measurement each would need to become falsifiable. Trace-dominance in comparative-evaluation races. Effort invested during encoding is proposed to lay down a deeper hysteresis trace that then carries a larger signal-share in later comparative-evaluation races. This is offered as one component of an effort-essential subset of the six biases, not a single-mechanism reduction: endowment under passive-ownership manipulations and sunk-cost under commitment-without-effort are well documented in conditions where the trace component cannot be operative. Commit-position in the response trajectory. Where a race commits in its output, and how the spread of that position is shaped by training, are themselves measurable. A probe on Qwen2.5-7B base versus its instruction-tuned counterpart (identical fine-tuning content) shows the base model with a 3.4× wider cross-condition spread in commit proportion, and a recognition-to-commit coupling (Pearson r = 0.528) that the instruct model loses (r = 0.104). Claims are graded by empirical readiness: established biases re-expressed in the framework’s vocabulary (parsimony, no new content); a candidate derivation with surplus content the native literature does not make (a predicted inverted-U over effort intensity; the commit-position spread×condition signature separating base from instruction-tuned substrates); and speculative extensions flagged as such. Appendix B engages the accumulator-model literature (drift-diffusion, leaky competing accumulator) so the proposal can be assessed on its own terms. Companion work: Behavioural Friction Theory (10.5281/zenodo.19462499); the substrate framework (10.5281/zenodo.20012654); matched friction under hysteresis (10.5281/zenodo.20059863). v2 update (July 2026). The preliminary, programmatic status of the paper is unchanged; this version tightens its theory and positioning. The earlier reading of commit-position as an optimal-stopping problem with an optimum near 1/e is withdrawn on theoretical grounds: on the race account commitment is obligatory at the output deadline rather than a threshold the trajectory crosses, so there is no stopping optimum for the position to sit at. The commit-position readout is repositioned against recent work that measures when a language model commits; the metric used here is described as a normalised, disclosure-adjacent proxy, and the paper's angle is narrowed to how training moves that position. An internal-commit-versus-public-disclosure distinction reconciles the no-optimal-stopping claim with the existence of learned disclosure policies. The effort-value section adds the nearest sunk-cost neighbour in language models at its scope carve-out. Finally, the joint-measurement commitment is sharpened: proposing that two readouts index one underlying process is a claim about selective influence, and because both readouts are currently read from the same source the joint study as first described would be single-method; a new prospective prediction specifies the double dissociation, with at least one readout obtained by a different method, that a genuine test would require. Builds on v1. v3 (August 2026) — self-containment revision. Key derivations are now stated in-paper; the disclosure-position proxy's identification gap is stated as a limitation; the substrate-general reading is advanced as a hypothesis within a resource-rational framing; small-sample results are reported at correspondingly reduced strength; editorial pass. Earlier versions remain in the version history.

Links and sources

Comments