The Appearance of Meaning: Context-Dependence and Semantic Competence in Transformer Architectures

Large language models routinely elicit interpretations of meaning and context-sensitivity from competent speakers. This paper does not argue that such models possess meaning proper. Instead, it presents a protocol-and-constraint framework that isolates an empir- ical explanandum–appearance of meaning (AoM)–understood as a competence profile exhibited under controlled contextual variation: (i) context-sensitive disambiguation, (ii) selective sensitivity to meaning-altering edits over meaning-preserving shams, and (iii) discourse-level constraint tracking beyond length-matched controls. The central ques- tion is mechanistic and philosophically diagnostic: in transformer language models, what internal variables causally control these meaning-like preference patterns when weights are held fixed? We test a Context-Primacy Thesis (CPT): that donor-directed changes in AoM-relevant preference margins can be induced by intervening on contextualized token-in-context states at characteristic depths, beyond sham baselines. Across GPT-2 and Qwen2.5 checkpoints, activation patching yields structured, depth-localized donor- directed effects with near-zero sham controls. A fixed-depth target-specificity stress test further indicates that, within a local positional neighborhood, causal leverage is often not uniquely localized to the ambiguous token position–constraining token-privileged imple- mentation glosses that treat context as merely a parameter applied to stable lexical carriers. The result is disciplined mechanistic constraint-setting for philosophical interpretations: an empirically anchored account of context-dependence in transformer computation that informs, without settling, disputes about semantic competence, reference, and normativity.