A remarkable result, due to Moore and Repullo (1988) is that any social choice function can be implemented as the unique equilibrium of a suitably constructed dynamic game (subgame perfect implementation). A recent literature has questioned whether common knowledge is a critical assumption for various implementation results (eg. Chung and Ely 2003 vis Nash implentation in undominated strategies). Directly related to subgame perfect implementation, Aghion et al. (2009) show that all extensive form mechanisms admit undesirable equilibria under an arbitrarily small p-belief perturbation from common knowledge, and that the desirable equilibrium in Moore-Repullo mechanisms fails to exist under such a perturbation. This paper test these findings in a laboratory experiment and finds that Moore-Repullo mechanisms exhibit this fragility in practice.