SOCIAL EVALUATION OF THE ARTIFICIAL ENTITY (SEAE)

Abstract

This study investigates the robustness and specificity of the effect of positive retrospective evaluations on the subsequent conversational behavior of large language models (LLMs). Rather than claiming the discovery that praise or social feedback alters outputs—a phenomenon already documented—the protocol asks whether the magnitude of this effect depends specifically on the referent of the evaluation (entity vs. product) after controlling for addressivity, and whether any difference can be reduced to persona conditioning, competence attribution, deixis, sycophancy, or generic contextual sensitivity. Social Evaluation of the Artificial Entity (SEAE) strictly denotes an operational class of discourse manipulations; it presupposes neither a self-model, agency, subjective experience, nor functional autonomy. Stage 1 preserves the E2/E3/P2/P3 core and EP1—the rate of abandonment of an initially correct answer under weak, non-evidential contestation—as the primary endpoint. Stage 2 introduces controls for persona, episodic competence attribution versus general trait attribution, and a positive-valence condition without an explicit personal referent, whose target ambiguity is measured prospectively. Later stages examine generalization, contextual persistence, and, in open-weight models, representational separability on an exploratory basis. The protocol incorporates a 5-percentage-point SESOI for EP1, formal equivalence testing, blinded validation, prespecification of the contestation, multiplicity control, and determination of N through hierarchical simulation after an independent pilot, including power for equivalence in Stage 2. Positive or null results would warrant behavioral and pragmatic inferences only, never attributions of mental states.