Beyond Prediction Error: A Decision-Theoretic Model of Resistance to Belief Updating

Agents often decline to adopt revisions their own inference supports. We propose that posterior computation and belief implementation are distinct operations: Bayesian inference produces candidate revisions, and a cost-sensitive decision — an endorsement gate — determines whether a revision becomes the operative belief, P(U = 1) = σ(β₀ + β_E·E − β_C·C), with E the posterior displacement (Bayesian surprise) and C the expected reorganization cost of implementation. The model adds a gating layer between inference and representational change without modifying inference. We derive the logistic implementation probability conditional on a specified random-utility representation, give a dynamic formulation with partial updating, and evaluate parameter recovery, estimator behavior, and rival discrimination across a reproducible simulation programme spanning identifiability, estimator pathology and correction, rival discrimination, adversarial severity, and design calibration, with four formal analyses. The theory’s confirmatory content is one preregistered discriminative program with two logically ordered stages, formalized as fixed-sequence gatekeeping. Stage one — state-trace coupling under selective single-channel cost manipulation — separates the gate from independent channelwise commitment thresholds: a reversed, two-dimensional association is fatal. Stage two — content-specificity in a matched-displacement, within-class two-belief design — separates a content-bearing operative state from content-general or content-conditioned threshold policies: a certified null belief-by-cost interaction is fatal, as is content-specific staleness without an implicit–downstream gap, or a positive pattern that retrieval probes expose as trace competition. An identity-conditioned threshold policy is shown to be observationally and psychofunctionally equivalent to the gate under a representationalist construal — underdetermination of ontology, not the victory of one vocabulary — so the empirical content rides on the shared structure. A severity-gated verdict table fixes in advance which outcomes are fatal, which supportive, and which merely silent. Every claim is flagged confirmatory or interpretive; alternatives are organized by a working taxonomy — empirical rivals, equivalent redescriptions, artifact families — whose placements are functional for the program, not verdicts on any account’s standing as a theory. The program opens with a pre-Stage-1 validation pilot whose success criteria are measurement properties, never effects, gated by preregistered traffic-light progression criteria. The register is held fixed throughout: a successful future experiment could support, at best, a shared, channel-coupled, content-specific, use-level operative constraint on belief use; the endorsement gate is the most parsimonious interpretation of that structure, and the choice of interpretive vocabulary is downstream of the result, not part of it.