The most important thing to know about grading in LawGrader: the assistant proposes, and you decide. A score is never final until you approve it.
What happens when you grade
When you grade a paper, the assistant reads it against your rubric and proposes a score for each criterion, with a short justification and margin comments tied to specific passages. It is doing the first read, the way a teaching assistant might, so you can react to something concrete instead of starting from a blank page.
Text students were required to copy
If your assignment includes wording students must reproduce word for word (a required format, a heading block, a signature line, a template), the assistant recognizes it in the paper and treats it as yours, not theirs. It will confirm the required wording is there and correctly reproduced, and tell a student when they left a bracketed placeholder unfilled. It will not critique that wording, and it will not count it as evidence of how well the student writes. Criteria about the student's writing are judged from the student's own text.
You do not have to set this up. It comes from comparing the paper to the assignment description, so pasting the required format into the description is all it takes.
How you stay in control
- AI scores are proposals. You adopt, adjust, or override any of them. Your number always wins. On a criterion scored in levels, each level shows the point range it covers, and you can fine-adjust the score in half points; the highlighted level always follows the score, and picking a level still awards its top value. A range that starts with a > excludes that number: a level reading >50–75 covers everything above 50 up to 75, while 50 itself belongs to the level below, whose range ends there. On a Checklist criterion the marks are the score: each item shows what it earned of its worth, and flipping an item's mark between met, partial, and missing recomputes the sum on the spot, so there is never an unexplained fraction.
- Every comment is reviewable. Keep the ones that help, dismiss the ones that do not.
- Every deduction's cost is editable. When a finding is right but priced wrong, open the criterion's Score Explanation and click its cost: raise it, lower it, or drop it to a no-cost note without deleting the comment. The ledger shows what you changed ("was −4"), and the assistant learns from the adjustment.
- Nothing reaches a student until you release it. Grading and releasing are separate steps on purpose.
- If part of a paper could not be read, you are told, not guessed at. Some criteria are judged from the look of the page rather than the words, and occasionally that read cannot happen: a submission has no PDF, or the read simply fails. Those criteria come back unscored and the paper shows as Incomplete instead of a number, because a score covering only part of your rubric would misrepresent the work in either direction. Score the missing criteria yourself and the grade completes normally.
How much help you want
You choose how much the assistant does, per assignment. It can simply read and surface observations, it can propose scores against your rubric, or it can draft student-facing feedback. You can dial that up or down at any time.
Calibration: teaching it your standards
Before you grade a whole stack, you can calibrate on a few papers you have already graded yourself. The assistant learns from where your scores differ from its proposals, so the rest of the batch lands closer to how you actually grade. This is optional, and it is one of the best ways to make the feedback feel like yours.
Bulk grading
Once you are confident, you can grade a whole set at once and then review the results, rather than opening each paper one at a time. You still confirm before anything is released.