How the tool is tested

Stress tests are not accuracy measurements. The synthetic descriptions are generated with the notation variants and their frequencies taken from about 15,700 published descriptions that were used to develop the tool. They show how robustly the tool reads the forms that occur in the literature, with the correct answer known for every description. They do not show how accurate the tool is on new articles: the accuracy of each version is measured only on test articles that were not used in its development.