This paper studys on the decision rule of the automatic grading system for Japanese dictation test which aims to certificate the reading and writing ability of Japanese Kanji and Hiragana. Different from general handwritten character recognition systems which allow to read incorrect (misspelled) characters, the grading system is required to discriminate miswritten characters more strictly from correct ones. This paper introduces the core processing stages of the system and focuses on the dicision rule of likelihood approach. Two threshold determination methods have been presented and comparatively evaluated. A grading accuracy of error rate less than 0.04% with rejection rate less than 70% is achieved in the performance test.