The concept of scaled achievement tests is discussed and a method of selecting those items of a test that form the most scalable (i.e., having the highest coefficient of reproducibility) subset is presented. Sometimes called a monotonic-deterministic model, this type of test assumes that the test items may be sequentially ordered. To determine the probability of obtaining a given coefficient of reproducibility, two statistics are required: the expected value of the coefficient; and the standard error of the expected value. The procedures used to eliminate the items that do not fit the model are described by example. Tests were constructed to measure two skills, elementary addition and elementary subtraction, and were administered to pupils in grades 2-6. The resulting data were statistically analyzed and a coefficient of reproducibility was calculated to determine how well the tests had been scaled. The tests were then administered to another sample. The results of the replication were much the same: internal consistency estimates of test reliability were almost identical for both tests. Most significantly, it was demonstrated that both of the hypothesized orderings of tasks were stable across samples. (Author/RC)