The most commonly cited early paper on the subject of automated grading in the CS context is Forsythe and Wirth's ``Automatic Grading Programs'' [13] from Stanford. In this paper, published in 1965, it is mentioned that ``grading programs have been used intermittently since 1961.'' The grading system that they present touches on many of the topics that have been brought up again and again in the ensuing years: tracking running time, storing student grades, terminating misbehaving submissions,3.1 security, and non-binary ``fuzzy'' grading. These themes take on charmingly antiquated meaning given that security concerns for Forsythe and Wirth focused primarily on hoping that students did not alter the grader library code on tape, and that students would actually use the pre-printed cards to set-up and tear-down the grading framework within their own submissions.
One of the quotes from this paper that blesses the growing field of CAA is this, ``We recommend grading programs to all who teach programming and numerical analysis to masses of students, but the prospective user should first carefully investigate the systems available to him.''
In terms of categorization of the system presented in this paper, Forsythe and Wirth reduce all need for estimation about robustness and generality, ``...it is relatively easy for us to write a separate grader for each problem, and furnish very detailed messages about each case.''