In Epistemology Naturalized, Quine suggests that we stop worrying about epistemic normativity, and just study the engineering problem of how to get into situations where we reliably form true beliefs. So, for example, we might do scientific studies on the reliability of witnesses of various kinds of events, under various conditions. Or, we might use informal mathematical arguments to show that all reasoning of a certain formal kind are truth-preserving.
Jaeguon Kim objects to this, by making the following claim: the notion of justification and epistemic normativity is necessary to make sense of the very idea of beliefs. Someone believes that P iff an ideal interpreter would assign them the belief that P. And such an ideal interpreter assigns them beliefs, by interpreting their utterances in such a way as to jointly maximize a) the simplicity of the interpreter's theory and b) the degree to which (on the whole) the subject comes out to have beliefs that are justified. Thus, it doesn't make sense to study the reliability with which someone forms true beliefs, while rejecting the notion of epistemic normativity.
However, I don't buy that in order to understand the notion of belief we must accept some kind of analysis of it into other terms. You might think: we are just trained in the practice of interpretation, like we are trained to recognize certain things as games. We don't do this by consciously reasoning about justification, and Davidson's maxims or any other thing that one might use to try to define the notion of belief. Maybe there aren't any informative necessary and sufficient conditions for having a belief that P, or the only conditions are extremely complicated and will only be discovered after years of work by linguists.
If this is right, the argument `Unless the notion of justification is coherent, there will be no informative analysis of what it takes to count as having a given belief! Therefore, the notion of justification is coherent.' looks pretty unconvincing. Maybe the notion of belief is primitive.
Note, that saying there need not be necessary and suffcient conditions does not mean that there cannot be interesting cognitive science done, breaking down our capacity to recognize when a subject S counts as standing in the belief relation to a proposition P into various components.
Consider the following little fish. It has (among other things) four sensors a, b, c and d. Whenever a and b are touched it says "I feel xish", whenever c and d are touched it says "I feel y ish" and whenever a and c are touched it says "I feel z ish". There are no informative necessary and sufficient conditions to be given within the fishes' language (assuming this is all his vocabulary that relates to these four sensors). And yet there is nice simple relationship between these three claims at the level of sub-personal cognitive processing.
Thus, if the naturalized epsitemologist prefers to take "believes that" as a primitive (by not attempting to define it in any other terms) this doesn't suggest any kind of defeatism about the power of cognitive science to explain our complex linguistic capacities by appeal to simple systems.
But maybe, I didn't need to say any of the stuff in the last three paragraphs, because, (usually?), the proponents of epistemic normativity think its irreducible. So, I should think, it's no worse to take there to be primitive facts about the believing relation, than about epistemic normativity.