We develop a methodology for testing computational hypotheses about neural functionality articulated in models at the systems level of description. In this approach, the first step is to attempt the construction of a model of the underlying brain system which is consistent with the known anatomy and physiology, but which is also able to exhibit functional properties consistent with a putative computational hypothesis. If this is successful, the second step consists of including additional known pathways into the model and testing the new models to see whether they show an improvement in functional performance (using appropriate performance metrics). A positive outcome is taken as evidence in support of the hypothesis. A final step is to construct 'control' models by including pathways that are not consistent with biological data. In this case a performance detriment is taken as support for the hypothesis. The methodology is applied to the basal ganglia, and builds on a previously published model of this system (Gurney et al 2001 Biol. Cybern. 84 401-23) which was based on the hypothesis that the basal ganglia perform action selection. The realistically constrained models show a selection benefit, while control models show a decrement in selection ability. These results, taken together, provide further validation of our selection hypothesis of basal ganglia function.