Students at grades four and five were administered a writing assessment that was developed to correspond to the California Learning Assessment System (CLAS) writing tasks at grade four. Teachers were trained to score the CLAS-like tasks according to the rubric developed by the State for CLAS. In addition, 164 students at three schools in the Riverside Unified School District, California, took part in the CLAS student level pilot in Spring, 1993. A generalizability study was conducted using a one facet, crossed design. The outcome was then used to conduct a decision study to determine how many tasks would be required to achieve various levels of student-level reliability. Comparisons of school level performance summaries were made between results from the CLAS assessment at grade four and results of the district's CLAS-like assessment. Results indicated that performance varied considerably both in terms of percents on the score categories as well as in mean scores. Local scorers on the CLAS-like tasks tended to place substantially more students at both extremes of the rubric than did the scorers of CLAS. In many cases, the rank ordering was markedly different. Results also indicated that at least five separate tasks scored and averaged would be required to achieve adequate levels of reliability. (Contains three tables of data.) (Author/RS)