With the contrast she made between WIMPs+tiredlight and Higgs+bigbang, I think she suggested that when both languages have equal predictive power, one should choose the language that is consistent with existing definitions. If a new language confuses definitions and creates the appearance of paradoxes, it is not a good language. This is drawn out in some of her other articles.

That only is really true in a proverbial vacuum - the choice will still affect the context within which these alternates exist. And as such, it will influence the perception of other measurements and the different options that will be explored to explain new data, so it really is not that arbitrary at all.