Two papers on classification, from Craig Cooley's dissertation. The Biometrika paper describes a simple, intuitive way to produce a classifier that is competitive with the best of today's modern, complex methods. The key to the classifier is turning Fisher's discriminant analysis from a normal-theory problem into a nonparametric problem while not falling prey to the curse of dimensionality. The paper in the Canadian Journal uses asymptotics to assist in prior elicitation.
A few more papers. For the Teaching Statistics paper, are you one of those folks who would never use an estimator that's so bad that it can assume values outside of the convex hull of the parameter space? If you've done survey work, estimated a proportion, and calculated a margin of error, chances are you've used one of these horrible estimators...