Vision-Language Models Assimilate Where Humans Contrast: A Cross-Architecture Signature of Contextual Computation

Which “cognitive biases” and “perceptual illusions” are necessary signatures of a bounded decision architecture, and which are mere artefacts? Behavioural data alone cannot say, because the latent decision variable must be inferred from aggregated choices. The per-token competing-routes (log-probability) substrate of a language model makes that latent observable, and can therefore act as an instrument that sorts a candidate phenomenon into phantom (an observer-aggregation artefact), architecture-forced (a real bounded-commitment signature), or substrate accident. This paper states that sorting criterion and demonstrates it in perception, using vision-language models (VLMs) as a no-retina visual substrate. With a gate-controlled, prior-orthogonalised forced-choice readout, classical inferential illusions recur as contextual modulation with a measurable commitment signature — but with a substrate-specific sign: VLMs modulate a target toward its surround (assimilation) where human perception modulates it away (contrast). The effect is a graded dose-response that replicates across three architecturally distinct vision encoders (a native dynamic-resolution ViT, a SigLIP encoder, and a CLIP encoder) and across two perceptual dimensions — size (Ebbinghaus, Delboeuf) and brightness (simultaneous-contrast) — with every one of nine dose-response slopes’ 95% confidence interval excluding zero. The opposite sign is strong evidence against the simplest account — that VLMs imitate human illusion reports absorbed from text — and is consistent with a systematically different contextual computation; whether that sign is forced by the architecture or learned from training-data statistics is left as the live open question, with the cross-corpus convergence (the same sign across three models trained on different data) weighing against a dataset-specific account. Companion papers in the series develop the cognitive and self-knowledge faces of the same “biases-as-architecture” claim, the social-friction worked cases, and the competing-routes measurement-model programme. Prepared for submission to Transactions on Machine Learning Research (TMLR). Data and code. The stimulus generators, evaluation notebooks, and per-token log-probability outputs are available from the author. v2 (August 2026) — attribution revision. The race reading is credited to its origin: treating two inputs as separate processes whose first finisher determines the response is the race model introduced by Raab (1962) to account for statistical facilitation, with Ratcliff (1978) as the single-process accumulation alternative. The limit is stated with it — this paper uses the architecture on a different substrate and does not test the race-model inequality that separates race from coactivation, so it adjudicates nothing between them. A companion previously cited as in preparation has been published and is now cited to its record. Editorial pass. Earlier versions remain in the version history.