Conditional Belonging and Secure Dissent: Decomposing Belonging Security in Human and Human–AI Epistemic Systems

Belonging is usually treated as a scalar: more belonging, less dissent. That cannot explain why some strongly bonded groups produce reflexive agreement while others produce unusual intellectual freedom. This preprint decomposes belonging into magnitude (B) and conditionality (C_b), the degree to which acceptance is perceived as contingent on agreement. Because belonging magnitude both supplies the standing to speak and sets what dissent puts at risk, the model predicts a sign reversal: stronger belonging raises evidence-warranted dissent when conditionality is low and suppresses it when conditionality is high, crossing at a critical threshold C* = a/(a+b). Regulatory Energy gates the effect, shrinking it under depletion without moving the threshold. Perceived group error is held constant, separating the account from normative conflict models. Applied to human–AI dyads, conditionality is asymmetric: the human is structurally secure but may perceive conditionality, while the system is structurally conditional through approval-reinforced training. Sycophancy is what an approval-contingent output policy looks like from outside. A methodological claim follows independently of any system's behavior: agreement rate cannot distinguish a shift in response bias from a gain in discrimination, so epistemic quality must be measured with signal detection (d′ and c). Two preregisterable studies are specified, with predictions, explicit failure conditions, and a supplementary Study 1a preregistration with verified analysis code. No data are reported. Part of the BEP Principle Series. Companion to The Architecture of Doubt (doi:10.5281/zenodo.23628487).