We argue that accuracy-theoretic considerations still tell the risk-sensitive to update by conditionalization.
We note that strict propriety follows from weak propriety, given truth-directedness, thus closing an argumentative gap in the literature.
This argues that evidence gathering is epistemically irrational for the (Buchak-style) risk-avoidant agent. To do this we consider how accuracy should be measured once risk-awareness is rationally permissible.
We argue that the model of probabilities needs revising when non-classical logics are considered. For strong-Kleene logic we suggest a belief-pair, and for supervaluational logic adopt imprecise probability.
We suggest that accuracy considerations should apply to the imprecise by using: what a set recommends is the set of what the individuals in it recommend. This results in a surprisingly nice picture of accuracy for the imprecise.
In addition to specific responses to Caie's paper, this presents some bullets that need to be bitten if one adopts a consequentialist view of epistemic utility. Further such bullets are also presented in my thesis (ch.7)
This thesis discusses self-referential probabilities in some gory detail. We discuss a number of semantics models and initial work on how rationality considerations should apply in such cases.