Rule-based spoken dialogue systems require a good regression testing framework if they are to be maintainable. We argue that there is a tension between two extreme positions when constructing the database of test examples. On the one hand, if the examples consist of input/output tuples representing many levels of internal processing, they are finegrained enough to catch most processing errors, but unstable under most system modifications. If the examples are pairs of user input and final system output, they are much more stable, but too coarse-grained to catch many errors. In either case, there are fairly severe difficulties in judging examples correctly. We claim that a good compromise can be reached by implementing a paraphrasing mechanism which maps internal semantic representations into surface forms, and carrying out regression testing using paraphrases of semantic forms rather than the semantic forms themselves. We describe an implementation of the idea using the Open Source Regulus toolkit, where paraphrases are produced using Regulus grammars compiled in generation mode. Paraphrases can also be used at runtime to produce confirmations. By compiling the paraphrase grammar a second time, as a recogniser, it is possible in a simple and natural way to guarantee that confirmations are always within system coverage.