LFG-DOP (Bod and Kaplan 1998, 2003) provides an appealing answer to the question of how probabilistic methods can be incorporated into linguistic theory. However, despite its attractions, the standard model of LFG-DOP suffers from serious problems of overgeneration, because (a) it is unable to define fragments of the right level of generality, and (b) it has no way of capturing the effect of anything except simple positive constraints. We show how the model can be extended to overcome these problems.