Although rule assessment approach for the study of children's problem solving is a promising direction, it presently has two drawbacks: the set of rules arising from the approach being not always well-grounded and complete, and the criterion of rule assignment proving statistically insufficient. This paper demonstrates that conceptualization of children's rules as production system is viable for clarifying researcher's rule generation process, and it demonstrates also that a modified version of latent class analysis proves to be a more suitable statistical model for evaluating children's task performance. The combination of production-system and latent class analysis concepts were applied to the data obtained by Noelting's juice problemamong 4th to 9th graders, and it was found that, in addition to the rules already proposed, the subtraction bug rule was adopted by the children.