In this paper, we devise a machine learning protocol to tackle a complex sociological task: to create a research sample from a few examples of interest, but in the absences of a clear definition of the target subset. As an example, we create a sample of organisational leaders starting from a list of nominees for the Bureau of the Intergovernmental Panel on Climate Change. The difficulty in this task lies in the impossibility to spell out the characteristics that define leadership in a complex and highly distributed organization like the IPCC. To bypass this lack of explicit definition, we use a series of techniques for anomaly detection to identify IPCC contributors with profiles similar to official Bureau nominees. We find that we can construct a precise (albeit implicit) model of IPCC leadership despite its social and political complexity, and that we can usefully use this model to expand our initial sample.