This paper examines Roma pupil segregation in the Czech Republic and explores why desegregation appears to succeed in some localities but not in others. Using a document-based realist evaluation, the study develops and refines context-mechanism-outcome (CMO) configurations for the city of Krnov, where a segregated primary school was closed in 2008 and Roma pupils were redistributed across the remaining schools. Drawing on publicly available policy analyses, reports, and media accounts, the paper identifies four linked mechanisms: dispersal that reduces school stigma, active municipal stewardship, a support infrastructure including tutoring and multi-agency cooperation, and school leadership that helps stabilise implementation. Secondary indicators suggest better-than-expected outcomes, including lower non-completion of primary education. The findings do not constitute causal proof, but they offer transferable principles and a “transfer test” for municipalities planning desegregation under different local conditions.