This work shows that the structure of constraint checkers can be exploited to dramatically accelerate sampling without sacrificing correctness; reminiscent of how conflict-driven clause learning transformed SAT solving by extracting and reusing information from failed search attempts. 6/