This may be obvious but, clearly there are (at least) two general types of errors: those caused by data external to the program and those caused by bugs in the program. For all inputs coming into the program from outside, such as user inputs and data coming over a network, the inputs must be completely checked -- always assume that they will be incorrect and are intended to crash your code -- be pleasantly surprised when they are not :-). Check for all bad inputs.
If the data are from inside the program, the assumption may be that they are good and if not, it was a bug. I suppose you can write unit tests on each routine to see that the methods that routine calls are with valid arguments (using mocks). I actually tried this in a Java program using JMockit; it was tedious, but it did catch a few "potential" bugs. I would love to say that I ALWAYS remember the input and output condition of every subroutine I write, but I just finished a new feature for a company project that involved 100+ routines and classes and I admit that I don't always remember which ones can return null and which always return empty arrays, for example, even though I try to write JavaDocs at the top of each routine. By mocking these routines, I can check that the caller always handles each case gracefully. -- http://mail.python.org/mailman/listinfo/python-list
