- This paper presents formal proof obligations for data refinement in the presence of unexpected crashes, notably due to a power failure. The work is part of our effort to construct a verified file system for flash memory. We apply the theory to one of the components in the flash file system, namely the erase block management layer. We show its functional correctness with respect to a high-level specification. We prove that the system can always recover from power loss to a desired state. We observe two simplifications that greatly reduce the proof effort for crashes in practice. Proofs are mechanized in the theorem prover KIV.