This is certainly possible. I've built a recipe creator where the operator will select their steps and can change a few items for each step that we allow. Then they can save the recipe and recall it based on serial number or other search criteria.
Underneath the hood we have some set steps in a template sequence file. When the recipe runs it is just copying steps over to a sequence file it creates on the fly and injecting the values the user set. Then it executes their recipe.
We are storing the recipes in JSON format but they could just as easily be XML or any other ASCII format.
The other advantage our architecture has is that we abstract the test algorithm away from the actual steps that do the pass/fail analysis. This makes our code super reusable, especially in a situation like you have where it's high-mix but similar family. In one case we used it on ultrasound probes and, like you said, there were just a few slight differences between the probes. We were able to add new probes almost instantaneously to our test system, as long as they could reuse the current algorithms.
If you are interested in discussing more via email or phone feel free to PM me as this can get a little crazy. Otherwise, you can ask specific questions here on the forums.
