That works well when your API is small and understood. But when you're exporting someone else's results (e.g., returning the result of a floating point operatio

That works well when your API is small and understood. But when you're exporting someone else's results (e.g., returning the result of a floating point operation,) your tests could start to become so exacting that they are more difficult to write and less reliable than the code you're testing itself. In that case, writing tests is basically writing an ad-hoc waterfall-style software specification. I tend to write tests when I'm not confident in the implementation or when the tests are trivial to write. At least, that's how I do it for new developments. If you're releasing code that people are relying on, you should have a whole suite of user tests and regression tests.