The New Standards Project conducted a pilot test of a series of performance-based assessment tasks in mathematics and English language arts at Grades 4 and 8 in the spring of 1993. This paper reports the results of a series of generalizability analyses conducted for a subset of the 1993 pilot study data in mathematics. Generalizability analyses for completely crossed designs of raters-by-tasks-by-pupils were conducted for a total of nine collections of mathematics tasks (samples of 30 to 63 students). The results of these analyses were used to estimate generalizability coefficients and standard errors of measurement for decisions studies using various combinations of number of raters and number of tasks. Consistent with results of previous analyses of performance-based assessment tasks, sampling variability due to tasks was found to be substantially larger than that due to raters. Implications for assessment designs are discussed. Includes nine tables and nine figures. (Contains 11 references.) (Author/SLD)