DRAT proofs have become the de facto standard for certifying SAT solvers’ results. State-of-the-art DRAT checkers are able to efficiently establish the unsatisfiability of a formula. However, DRAT checking requires unit propagation, and so it is computationally non-trivial. Due to design decisions in the development of early DRAT checkers, the class of proofs accepted by state-of-the-art DRAT checkers differs from the class of proofs accepted by the original definition. In this paper, we formalize the operational definition of DRAT proofs, and discuss practical implications of this difference for generating as well as checking DRAT proofs. We also show that these theoretical differences have the potential to affect whether some proofs generated in practice by SAT solvers are correct or not.

Inprocessing techniques have become one of the most promising advancements in SAT solving over the last decade. Some inprocessing techniques modify a propositional formula in non model-perserving ways. These operations are very problematic when Craig inter- polants must be extracted: existing methods take resolution proofs as an input, but these inferences require stronger proof systems; state-of-the-art solvers generate DRAT proofs. We present the first method to transform DRAT proofs into resolution-like proofs by elim- inating satisfiability-preserving RAT inferences. This solves the problem of extracting interpolants from DRAT proofs.