UCPOP is systematic: it will never visit the same partial plan twice while searching. Cassandra, as described in this paper, is not systematic; it may visit some partial plans in the search space more than once. Consider again the plan to disarm a bomb that we discussed in Section 4.1. In this plan, there are two different ways of establishing the goal to disarm the bomb: by dunking package1, and by dunking package2. Cassandra can initially choose either way of establishing the goal, leading in each case to the introduction of a contingency and the necessity of replanning to achieve the goal in the other contingency. Both search paths arrive at the same final plan, so the search is not systematic.
Cassandra could be made systematic by insisting on handling the contingencies only in a certain order, the search path that uses the other order being treated as a dead end. However, this extension has not been added as there is currently some debate as to the desirability of systematicity. For example, Langley [1992] argues that a non-systematic search method, iterative sampling, is often better than a systematic method, depth-first search, for problems which have multiple solutions and deep solution paths. Peot and Smith [1992] observe that the performance of a non-systematic version of was better than that of the original systematic version. They ascribed this behavior to the fact that exploring duplicate plans consumed less overhead than did ensuring systematicity.