Your boss has demanded that you develop a sorting algorithm to improve the performance of your company's application. However, having written the application, you know that you're unlikely to be able to make it significantly faster. Not wanting to disappoint your boss, you've decided to develop a new algorithm that works even better than *sort on certain sets of data. Of course, you can't make it obvious that the algorithm only works on some cases, so you want to make it obscure as possible.
The objective of this contest is to write a sorting routine in the language of your choice that performs better on certain sets of data than others, with repeatable results. The more specific the classification that determines speed, the better. The algorithm must do sorting of some kind, so an algorithm that depends on the data already being completely sorted (as in, an algorithm that does nothing), or an algorithm that depends on the data being completely sorted in reverse, are both invalid. The sorting algorithm must correctly sort any set of data.
After presenting your routine, please include an explanation of why it only works on certain sets of data, and include test runs on at least one set of good (fast) data and one set of bad (slow) data. The point here is to be able to prove to your boss that you've stumbled upon a better way to sort, so more test data is better. Of course, you're only going to show your boss the test results from the good data, so the flaw in the required testing data can't be too obvious. If applicable to your language, please show that your algorithm is faster than your language's built-in sorting algorithm.
For example, one might submit an insertion sort algorithm, with the good data being data that is already nearly sorted, and bad data being completely random data, since insertion sort approaches O(n) on nearly-sorted data. However, this isn't very good, since my boss would probably notice that all of the testing data is nearly sorted to begin with.
This is a popularity-contest, so the answer with the most votes after 7 days (May 21) wins.
If no one beats me to it, I'd like to submit a community wiki answer that takes advantage of uniformly distributed data sets.