And then we learned about what happens when you started giving the models input that were outside the boundary conditions… suddenly they just didn’t look so great.

One of the things about modeling we did learn was that it was best to use input parameters that had straightforward effects on the output… that is, the physics of the process was reasonably well understood. When an input parameter began to show a counterintuitive relationship to the output it was a great big red danger-flag. We found that it usually meant that troublesome parameter was actually a proxy for something else, and that we’d better find out what that something else was if you wanted to extend the model.

It strikes me that far too many climate modelers have fallen into the trap of believing they don’t have to understand the differences between driving parameters and proxy parameters. The computers give you the correlations and that’s all you need, even if that individual correlation doesn’t make a lot of sense. They’ve forgotten that one thing that computers are really, really good at is to make huge numbers of really stupid calculations in a short period of time. And some of those stupid calculations, while fundamentally wrong, will result in answers that appear to be correct. Those who put great faith in overly complex models usually get smacked by boundary condition assumptions that make those models appear to work.