[DougWebb]

Several years ago I wrote a blog post expressing a related point, contrasting software design to construction. Here's a part:

Let me give an example: if you're designing a bridge, you can draw blueprints on paper which shows girders. The girders are described by giving their dimensions (accurate to 1/16th of an inch, say) and the particular alloy the girder is made from. This is sufficient to accurately model how that girder will behave under all kinds of different stress loads which is important for ensuring the bridge will be safe, and also to model how the girders will fit together like a puzzle which is important for allowing the steelworkers to build the bridge correctly, on-time, and on-budget.

The key to all of this is the fact that you don't need to create a real girder in order to test the design and make sure it's correct. A few easily described properties of the girder are sufficient; it doesn't matter where every atom goes, it doesn't matter if the surface isn't perfectly uniform, it doesn't matter if there is some rust, etc. Lots of the details just don't matter at design time, and most of them don't matter at construction time either.

Software just doesn't work this way. Software development languages are extremely detail-sensitive: get one letter wrong, one punctuation character in the wrong place or left out, and the software won't work right. There is no way to accurately model something this sensitive to detail without building it first, and if you have to build it first you lose the biggest benefit of doing design up-front: the ability to test and iterate on your design cheaply before committing to a full build of it.