Certifying Model Checkers

Model Checking is an algorithmic technique to determine whether a temporal property holds of a program. For linear time properties, a model checker produces a counterexample computation if the check fails. This computation acts as a “certificate” of failure, as it can be checked easily and independently of the model checker by simulating it on the program. On the other hand, no such certificate is produced if the check succeeds. In this paper, we show how this asymmetry can be eliminated with a certifying model checker. The key idea is that, with some extra bookkeeping, a model checker can produce a deductive proof on either success or failure. This proof acts as a certificate of the result, as it can be checked mechanically by simple, non-fixpoint methods that are independent of the model checker. We develop a deductive proof system for verifying branching time properties expressed in the mu-calculus, and show how to generate a proof in this system from a model checking run. Proofs for linear time properties form a special case. A model checker that generates proofs can be used for many interesting applications, such as better ways of exploring errors in a program, and a tight integration of model checking with automated theorem proving.