Beyond Social Performance: Evidence Modes and Validation Obligations for Other-Agent Models in LLM Systems

Large language model systems adapt their tone, infer likely intentions, sustain conversational roles, and succeed at strategic social tasks. Such behavior is often interpreted as evidence for a theory of mind or an internal model of another agent. Yet socially appropriate performance underdetermines its mechanism: the same output can arise from memorized regularities, generic social policies, task-format shortcuts, mental-state descriptions that never affect action, or a state that is genuinely used in prediction and planning. This paper proposes criteria and experiments; it reports no new experiment. Its primary contribution is a two-axis framework that separates evidence mode—behavior, verbalized state, decodability, component intervention, or content-targeted intervention—from the validation obligations required for a warranted attribution, including matching quality, agent indexing, content validity, system-boundary validity, necessity, sufficiency, mediation, selectivity, and cross-context generalization. We distinguish an other-agent model, which represents another agent's task-relevant state, from a self-model, which represents the system's own state. We then present two bounded operational extensions. First, we separate established textual counterfactual, perturbation, and matched-contrast designs from a structured intervention that changes a target agent's observation history while counterbalancing the evaluated system's direct evidence and action affordances. Second, we propose a falsifiable social-self candidate in which socially indexed evaluation changes a persistent, self-indexed state, and test it against episodic retrieval, generic reward conditioning, partner-specific policy, role-conditioned policy, and reputation-database nulls. The literature reviewed here supports the need for these distinctions but does not establish targeted causal use of a specific other-agent representation or the proposed social-self topology. The aim is not a new binary verdict about whether an LLM “has” theory of mind, but an auditable account of what is represented, how it is revised, where it is used, and which consequences persist.