This note examines the role of Weldon's classical dice data in illustrating the problems of model selection.New, more realistic models are discussed; simple Poisson trials models, with at most two unknown parameters, are fitted.An application to process-oriented inspection sampling for a multiple channel production process is suggested.