The AI marks. Our code adds up. The grader marks four dimensions, each on its own, against the rubric you can read below. Fixed weights in code work out the total, and the pass mark is 70 for everyone.
Fluent writing is not enough. One dimension on every rubric is a gate. An answer that commits to nothing cannot pass, however polished it reads.
Every part is checked, and quoted. The grader lists each part of the task as present, thin, wrong or missing, quoting the words it judged and the fact it checked them against. A quote that is not in your answer does not count. One thin part caps the total at 84. A wrong or missing part caps it at 74.
Every grade names its grader. The grader runs a fixed model version at its lowest-randomness setting. Each change to it is a numbered release, stamped on every graded record and shown on your proof page.
What it does not do. It cannot tell who wrote the answer. CompassStu grades the work; it does not verify who wrote it.