This squib presents a simple abstract learning problem on which the Gradual Learning Algorithm (GLA; Boersma 1998, Boersma and Hayes 2001) fails. It then discusses the relationship of the GLA to the provably convergent Perceptron learning algorithm (Rosenblatt 1958), and shows that the learning problem is solved by a combination of the Perceptron update rule with Harmonic Grammar (Smolensky and Legendre 2006).