The Architecture of Doubt: A Testable Model of Confidence–Evidence Decoupling Under Social and Cognitive Constraints

Subjective confidence and evidential support can diverge, and social context is known to shape confidence. This preprint states a narrower, falsifiable Confidence–Evidence Decoupling (CED) hypothesis: restricted deliberative resources and credible relational costs of disagreement interact to change how strongly confidence responds to diagnostic counterevidence. Building on the Belonging–Energy–Prediction (BEP) framework, the paper specifies a trial-level confidence-updating model with a fixed choice referent, incentive-compatible elicitation, and a randomized 2×2 design in which the post-evidence confidence report is private in every cell. The main formal result is that the sign of the focal interaction is not fixed by the verbal hypothesis. Multiplicative gating makes the three-way coefficient necessarily positive. Threshold gating, as proposed by BEP's Courage operator, makes it negative exactly when single-constraint conditions remain above threshold, while a log-scale contrast is negative in every regime. Two co-primary estimands therefore separate gating architecture from operating regime. The paper lists competing normative and descriptive models, the predictions of each, and explicit failure criteria. It reports no empirical data; the figure and numerical checks are analytic, and the script that reproduces them is included. Part of the BEP Principle Series. Extends The BEP Field Theory of Decision (doi:10.5281/zenodo.17759583) from update-versus-conform dynamics to the evidence-to-confidence mapping.