I think your objection misses the fundamental point of the paper which is that blindly applying formulae (in this case Bayesian ones) without considering the part these computations play in the whole scientific process leads to bad science. Specifically, it leads to assuming that the correct model is already among those being considered. The authors say this is often not the case and assuming so often leads to stagnation in the relevant discipline.

This isn't an article that hands the readers a set of rote instructions, it's a warning that rote application is bad science because good science doesn't merely compare existing models with the data, good science proposes new and better models. This is fundamentally an article on the philosophy of science so leaving out the philosophy would make the article pointless.

In general, people who are good at symbolic manipulation are often in search of a methodology that allows them to only do symbolic manipulation and relieves them of the difficult task of doing semantics as well. The article is saying that there's no free lunch here - we have to do the semantics as well, we have to understand why models don't perform well and come up with the creative insights necessary to replace them with new ones.