A program is a theory. A theory can be evaluated: firstly, does it describe/predict the data (i.e. does it work)? Secondly, is it parsimonious? But the key is

A program is a theory. A theory can be evaluated: firstly, does it describe/predict the data (i.e. does it work)? Secondly, is it parsimonious? But the key issue with a program-as-theory is that you must understand the theory to understand the program. Because of this, it may be much more effective to shoehorn a problem into an ill-fitting theory, if that theory is already well-known and well-understood. A new, unfamiliar and strange theory must be REAL GOOD - and that improvement must also be valuable in the applied context - for it to be worth using. If so, it becomes a new standard theory.