Solving crossword puzzles involves a memory-based decision process using two simultaneous cue types: semantic and orthographic. To understand this type of constrained decision process, we have developed a cognitive model of these processes that we demonstrate can be trained to solve puzzle questions as well or better than most humans. The model enables us to test how humans perform the search process, by examining alternative accounts of knowledge access. We conclude that the most reasonable explanation for human crossword play is a dual-route process, by which each of two routes are searched independently and consecutively, with the fluency of the semantic route being the most important factor distinguishing novice and expert players. We conclude by discussing general implications for expertise as a recognitional decision process.