Well, one important thing to avoid is the type of argument that requires massive exhaustive search through a vast space to find a flaw. It can be perhaps escaped by declaring such arguments void unless the full exhaustive search has been performed by a computer.

I had some thoughts for the models. An argument based on the imperfect model of real world should outline imperfections of the model and then propagate the imperfections along with the chain of reasoning, as to provide upper bound on the final error. It is often done with computer simulations of e.g. atmosphere, where you can e.g. run simulation a lot of times with different values that are within the error range and look at the spread of the outcome.