In an excellent discussion on the CriticalChain Yahoo Group, Larry Leach mentions a point I hadn't put together before. "If you do not have a statistical basis for predicting the project end date, you have no
basis at all."
Essentially, even with a "unique" project, the elements of the project (the tasks) cannot be unique. Otherwise, the entire project and task duration estimates are pure guesses -- and it is worthless to assume you can even complete the project. Said another way, through experience and measurement, we know approximately how long it takes to do our work - this knowledge is the root of Leach's statistical basis.
It's a revelation to me.