Fine-Grained Evidence

Bayesian conditionalization is rigid: learning E fixes p(E) at 1 while preserving probabilities conditional on E. Non-rigid update is preferable when, in the course of learning that E is true, we change our views about how—by way of which truthmakers ϵ. A Jeffrey-style generalization of Bayes—active conditioning—is developed which gives learning events a handle on p(ϵ|E) and p(E) both. E brings a truthmaker-incorporating “probasition” to the table, rather than simply an intension. Confirmation relations go hyperintensional as a result. Eis true in the same worlds may not license the same updates, if their truth flows from different sources.