The puzzle highlights the inability to model a specific kind of defeasibility of evidence in a Bayesian framework. But the puzzle is of much wider interest: it arises not just for Bayesians, and not just for formal modelers, but for any theory which adopts three very plausible principles: the rigidity, symmetry, and defeasibility of evidence. In this paper I first present Weisberg's puzzle as originally developed, in the context of Bayesian epistemology. Then I show how the problem arises for normal epistemology as well. 1 I offer a solution for the normal puzzle and then carry over that solution to the formal side. I end up by suggesting that taking the puzzle seriously means facing up to the possibility that we cannot give an informative, general characterization of epistemic justification, and that we should instead rest content with a kind of particularism. 2 BAYESIANISM AND DEFEATERS Bayesianism in epistemology takes degrees of belief, or credences, as the doxastic attitude to be modelled. Credences are formally represented by a function, , which takes propositions as arguments and delivers real numbers as values. Traditional Bayesianism imposes two normative constraints on credences: at any given time, rational credences satisfy the probability calculus (call this constraint "Probabilism"), and, upon receiving evidence , rational credences evolve by conditionalization on (call this constraint "Conditionalization"). Credences obey the probability calculus if and only if all propositions are assigned credences between 1 and 0, logical truths are assigned credence 1, and the credence assigned to a disjunction whose disjuncts are logically incompatible is the sum of the credences assigned to each disjunct. The conditional probability of a proposition given another proposition is defined as follows: (|) = (∧) () (provided that () > 0). In this traditional Bayesian framework, which propositions a subject has as evidence at a given time is taken as an exogenous input, on which the framework is silent. Once that input is set, Conditionalization is the norm according to which, for any proposition , 54