There was originally a separation between 'TransactionValid' and 'TransactionInvalid' exceptions. We removed this before milestone 2 because at the time it seemed like exception handlers could discriminate between any number of types and decide on it's own what should cause the 'chunk to fail'. However, recent changes to the architecture have caused us to rethink that. Exceptions that happen on reading really need to be treated differently from those encountered while writing, generally because of transactions. Therefore, there really is a domain concept of 'read failures' and 'write failures' and policies that should know how to handle them.