Much previously published material for estimating the reliability of classification has been based on the assumption that a test consists of a known number of equally weighted items. The test score is the number of those items answered correctly. These methods cannot be used with classifications based on weighted composite scores, especially if the composite includes essay scores. This paper presents a modification which will make it possible to apply these methods to composite scores. The proposed method is based on a normal model (with variance stabilizing transformation) for the conditional observed score distribution. The effective test length of the composite is determined from its true-score variance, estimated by Kristof's method or by Gilmer and Feldt's method. (Author/DWH)