Exploring questions of representativeness, balance and comparability is essential to tailoring corpus design and compilation to research goals, and to ensuring the validity of research results. This is especially true when the target population of texts under examination is very large and transcends a restricted area of specialization and/or covers multiple genres, as in the case of texts translated in institutional settings. This paper describes the multilayered sequential approach to corpus building applied in a comparative study on legal translation in three of these settings. The approach is based on a full mapping and categorization of institutional texts from a legal perspective; it applies an innovative combination of stratified sampling techniques integrating quantitative and qualitative criteria adapted to the research aims. The resulting corpora, categorization matrix and selection records, together with the methodological detail provided, can be useful for building other multi-genre corpora in translation studies and further afield.