We experimentally test some qualitative predictions of the Kreps-Wilson (1982b) model of reputation building in a version of a borrower-lender game first used by Camerer-Weigelt (1988). Our results do not differ substantially from those of Camerer-Weigelt--when we use the same parameters they used. However, we find that the theory fails to account for observed behavioral variations to parameter changes. There is a systematic response to changes in the payoff function of the borrowers, but this response is not in the direction predicted by the theory. This cannot be reconciled with the theory by an appeal to "homemade" priors of the type specified by Camerer-Weigelt. Furthermore, the estimated magnitude of "homemade" priors in the Camerer-Weigelt experiment is inconsistent with estimates derived from behavior observed in our own experiment.