The probability of classi"er errors in automated grading of fruits is much greater than in traditional well-de"ned and highly separated, static classi"cation tasks. Presently, operators of conventional sizers and colour sorters adjust the class boundaries manually based on observations of obvious misclassi"cation trends in the packed fruit, with the goal of minimizing the classi"cation errors. However, the new sorting machines utilize many features to reach the grade decision. A human operator is unable to control the multitude of parameters under control. Estimating the between-class discriminant functions requires estimation of the a priori class probabilities (&priors') and the class-conditional probability densities. The time-varying nature of the priors and the probability densities result in unsatisfactory classi"er performance. To solve these problems, an adaptive grading approach by &prototype populations' is proposed. The produced stream is classi"ed into a discrete number of prototype streams or populations by a global &population classi"er'. For each unique prototype population a separate, optimal &grade classi"er' is designed for sorting individual fruits. The global &population classi"er' utilizes a "nite-length stack of features continuously updated from the most recently sorted produce. The statistical attributes of the features sample in the stack are analysed to determine which produce population is currently passing through the system. When the population classi"er determines that the stack contents have originated from a di!erent prototype population, it changes the active &grade classi"er' to the most appropriate one for the current fruit population. An example of simulated adaptive versus conventional train-once, sort-many, grading is presented on data sets obtained from a system to sort dates by machine vision. The example demonstrates that adaptive grading by prototype populations yields lower misclassi"cation rates in comparison to conventional sorting.