By the way, I see that for count-graphs, the ultimate benchmark, using both call-based and continuation-based was a 26% speedup over just call-based. Definitely not too bad. You never said what you thought the problem was with the cases where both ran slower. Any guesses?