We investigate the empirical applicability of several bounds (a number of which are new) on the true error rate of learned classifiers which hold whenever the examples are chosen independently at random from a fixed distribution. The collection of tricks we use includes: