One thing that we almost NEVER do is with models is force them to interact with each other. This means that we are completely blind to an entire aspect of isolated models: namely that inputs may not have the structure we expect, or that their outputs don't actually have the anticipated affect on downstream models. Most current model validation is purely speculative or based only on a finite dataset. If you take two models that you think effectively represent how the world works and put them together and suddenly they no longer work, then you reveal a huge gap in our knowledge. In a sense this is the ultimate form of model validation and until you do it any interpretation you make about how models might or might not work together is complete bullshit.