I comment previously on a draft version of this paper. In this ‘final’ version the authors have removed the wrapper function which converted predicates (:: STM a -> STM Bool) into actions that raise exceptions when the invariant doesn’t hold (:: STM a -> STM ()) leaving the interface a little more consistent.

They have extended the STM monad to use phantom types to restrict the operations an STM action can perform (ReadOnly or Full) to help restrict the side effects of invariants. Another addition is support for invariants over pairs of states: the state before the transaction began, and the state about to be committed. This is accomplished with a new primitive: old :: STM a -> STM a which executes an STM action in the state as it was when the current transaction began. Finally, there is some new discussion of drawing on the work in databases where a distinction is made between assertions, which detect error conditions, and triggers, which are part of the programmes logic.

A good paper made even better!