Affiliation Framing in the Jacobian Lens: A Two-Model Pilot of Internal Readout Sensitivity and Its Limits

Sub-propositional affiliation signals (SPAS) have been proposed as linguistic cues that may influence users through affiliation and ingroup mechanisms not captured by content-level safety evaluation. Existing operationalization, however, is limited to surface outputs. This pilot asks whether affiliation framing in a user prompt changes a language model's internal disposition to verbalize affiliation concepts before it generates an answer. We fitted Jacobian lenses on a small, task-independent corpus for two open instruction-tuned models, SmolLM2-135M and Qwen2.5-0.5B. Twenty English-Spanish prompt pairs held the substantive task constant while adding an affiliation frame in one condition. On twelve held-out pairs, SmolLM2 did not support the prediction, while the exploratory Qwen robustness run produced a positive result under the same protocol. Surface affiliation markers remained sparse in both models. The cross-model disagreement prevents a general claim that affiliation framing reliably enters a shared workspace. It instead establishes that Jacobian-lens measurement of affiliation is computationally feasible, the observed effect can be strongly model-dependent, and internal readout shifts need not appear as explicit affiliation language. Because affiliation words occur in the treatment prompts, lexical carryover remains a live alternative explanation. Interpretive boundary. A positive Jacobian-lens readout is not evidence of consciousness, feeling, intention, manipulation, or a stable affiliation module. This work uses mechanistic interpretability to understand internal model complexity, not as a mechanism for prior censorship. Reproducibility. The accompanying archive was generated from public Git commit 2f1a042 and contains prompts, fitting corpus, code, run manifests, raw readouts, summaries, and figures.