I've sometimes used the inverse of this (reading the tests to understand the intended usage), but had never thought of it as an exploratory technique. Further consideration suggests that these experimental test files might be useful to the author, in at least two cases:

1. when reporting a code bug or patch.
2. when reporting a documentation bug or patch.

The first of these is recommended practice, but the second might also be useful. When receiving a documentation patch or question, it's sometimes hard to tell why the person asking the question doesn't understand my <sarcasm>perfectly clear and understandable</sarcasm> documentation. I can see the test file containing your experiments being very useful in making assumptions and misconceptions slightly clearer.

While some might argue that this isn't the best way to learn an interface, I think you have added another tool to my learning toolbox.