Brief summary saved:

Basically the aim is to give the user (domain author) a good example of the different kinds of stories their domain can generate, since the actual number of different stories is intractable.  To do this, we can first build a plan graph out of the domain in question... (recall, a plan graph ignores delete lists... so it depicts all the ways to achieve "all the goals" but doesn't account for the fact that goals could be undone)...  Once we have the plan graph, find all the "relaxed" solutions therein. Then, using some distance metric e.g. Jaccard distance (pull from existing research on doing this with narrative plans), cluster the solutions into k buckets based on their similarities. The user could provide k or we could be fancy and have it automatically find the best value for k. Once grouped into buckets we just need to display the information to the user, e.g. by showing a prototypical example of a plan in each bucket