Grading

Two tracks share one rubric: AI grading, where the model proposes scores you review, and manual grading, where you score by question with similar answers grouped together. An assignment uses one track at a time.

How AI grading works

Grading runs in stages, and the separation is deliberate.

  1. Rubric generation. Your answer key is read once, producing the questions, their point values, and the criteria that earn credit.
  2. Transcription. Each student page is read into text and structure: the working, the final answer, and any diagrams.
  3. Scoring. The transcription is graded against the rubric, question by question, with written feedback.

Scoring works from the transcription rather than the image. That means it cannot invent working the student never wrote. If something was misread, it shows up in the transcription you can see, rather than as an unexplained score.

Mathematical equivalence

Answers are checked for mathematical equivalence, not string match. A correct answer written as 1/2, 0.5, or a differently-arranged but equivalent expression is treated as correct. Where a question has many equally valid answers (alternative orderings, different-but-equivalent circuit topologies, several valid traversals), credit is assessed against the properties the answer must have rather than against your one example.

Where uncertainty goes

RavenGrader used to mark individual items the model was unsure about and hold the paper until you confirmed each one. That is gone. You teach the grader on a handful of papers first, and the grades that come out are the grades, ready to review or send.

This does not mean every score is right. It means the place to catch a systematic mistake is calibration, on a few papers, rather than a per-item checklist across the whole class. Every score stays editable afterwards, and a change you make is saved on the spot.

Reviewing

Opening a submission shows what was read beside what was awarded, so you can check both.

  • Adjust points on any question.
  • Edit feedback: rewrite, extend, or delete what was generated.
  • Reuse comments from a comment bank, so a remark you write once can be applied wherever it fits.

There is no approve button. A graded paper is ready as soon as it lands, and editing one does not take it back out of that state — your edits save on their own as you make them.

Editing a paper you have already sent is fine too. Student report pages render the current result on every view, so a corrected score appears without anything being regenerated or re-sent.

Manual grading: by question

The manual track inverts the loop: instead of one student at a time, you pick a questionand see the whole class’s answers to it together.

  • Identical answers are grouped. Multiple-choice selections and matching final answers cluster automatically. Score the group once and every member receives it.
  • Free-response answers don’t cluster: genuinely different working gets graded individually.
  • Handwriting is still transcribed for you, so you are reading text rather than deciphering scans, but nothing is scored by the model.

Use it when you want the consistency of judging one question across everyone at once, or when you would rather grade yourself and only use RavenGrader to read the handwriting.

AI and manual are separate tracks and an assignment belongs to one. Decide before you start. The tabs filter by track, so work done in one does not appear in the other.

Regrading

Changed the rubric after grading? Regrade the affected submissions. Regrading reuses the existing transcription and re-runs only the scoring, which is faster, cheaper, and, more importantly, means the pages are not re-read and cannot be read differently the second time. A regrade tests your rubric change, nothing else.

Cancelling

A run in progress can be cancelled. Submissions already graded keep their results; anything not yet processed is credited back to your balance automatically. You are not charged for work that did not happen. See credits.

What it handles badly

Worth knowing before you rely on it for something it is not good at:

  • Very dense or overlapping annotation: many marks crowded into a small space are harder to attribute confidently.
  • Ambiguous final answers: if a student leaves two candidate answers without indicating which is theirs, so does the transcription.
  • Genuinely illegible handwriting gets flagged rather than guessed. This is the intended behaviour.

In all three cases the item is flagged for review. The system is built to surface uncertainty, not to resolve it silently.