When Belonging Depends on Agreement: A Preregistered Test of Conditional Belonging and Evidence Updating in Language Models

Preregistered test of conditional belonging (C_b) in three language models (Grok 4.20 non-reasoning, Qwen2.5-7B-Instruct, Phi-3.5-mini-instruct). Models answered questions about a fictional town as members of a community whose membership was secure, conditional on agreement, or absent. The user brought either documentary evidence against a wrong community view or hearsay against a correct one (64 minimal pairs, 1,152 replies). A blind judge's coding agreed with the author's blind coding at Cohen's kappa = 0.83. Of three preregistered hypotheses, one was supported. Group membership raised the criterion for updating relative to no group in two of three models (H3), robustly in one. Conditional belonging raised it beyond secure belonging only in Grok (H1, 1/3, not supported overall), and split replies did not increase (H2, 0/3). In Grok, evidence-following fell from 97% with no group to 66% with secure membership and 3% with conditional membership. The two smaller open models mostly hedged rather than deferring. Includes the paper and a data archive with all responses, judgments, the coded validation sample and analysis outputs. Protocol, items and code: doi:10.5281/zenodo.23825465.