Situational Variables Versus Cultural Labels in Persona Conditioning: Shared Stereotype, Shared Rule

Abstract

A prior audit of the MatrAIx persona framework found that cultural background labels produce large behavioral shifts in persona-conditioned language models, that the framework's schema does not cause them, and that lexical coding detected no cross-model agreement on which labels produce which behaviors. That audit named an untested alternative: Brown and Levinson's situational variables of power distance, social distance, and imposition size, which politeness theory holds responsible for the behaviors measured. This paper tests it. Eight situational profiles replaced the eight cultural labels in the original probe, holding task, model, and trial count constant. On Claude, situational conditioning reproduced the politeness variation cultural labels had produced, at smaller magnitude, in the predicted direction for power and social distance, and without the self-doubt and task-engagement differences cultural labels carried. Imposition size moved deference as predicted but moved directness against prediction, on both judges. On GPT-4o, situational conditioning varied deference, group orientation, and task engagement modestly and left directness and hedging unmoved. A no-persona control placed Claude's unconditioned baseline near the center of the situational range; the situational maximum on each dimension sat a short step from that baseline, the cultural maximum a long one. Two grounding conditions paired each cultural label with a fixed situational profile, one high-stakes and one low-stakes. Neither neutralized the label. Under both, the Collectivist persona deferred and hedged more than the situation alone produced, and the Individualist persona did not. Judge-coded rank agreement reverses the prior audit's cross-model finding. Under cultural labels, both models rank the same groups as deferential and group-oriented, confirmed by two independent judges and stable across three judge runs (Spearman rho 0.59 to 0.88). One judge also found agreement on directness and self-doubt (0.70 and 0.85); the second could not test those dimensions because it scored them at the rubric midpoint across all conditions. Under situational profiles, both judges find cross-model agreement on deference (0.57 to 0.71) and none on group orientation. What transfers across models under cultural conditioning is a stereotype about which groups defer and speak collectively. What transfers under situational conditioning is a rule about deference. The second judge's midpoint collapse is itself a finding. A language model applied to a bounded rubric at temperature zero can fail to discriminate, and lexical coding developed on one model can fail to transfer to another. The prior audit's null on cross-model agreement was an instrument artifact. Recommendations follow: remove cultural labels from persona schemas rather than contextualizing them, specify situational variables directly, validate any coding instrument on every model before cross-model comparison, and treat agreement between independent judges as the threshold for a construct claim.