ABSTRACT. Based on our earlier formalization of conflict-driven clause learning (CDCL) in Isabelle/HOL, we refine the CDCL calculus to add a crucial optimization: two watched literals. We formalize the data structure and the invariants. Then we refine the calculus to obtain an executable SAT solver. Through a chain of refinements carried out using the Isabelle Refinement Framework, we target Imperative HOL and extract imperative Standard ML code. Although our solver is not competitive with the state of the art, it offers acceptable performance for some applications, and heuristics can be added to improve it further.