Larp Detected; Assignment Rejected

Mohak Subramanya Murthy — Sun, 15 Nov 2026 07:24:28 +0000

There is a beautiful moment after submitting homework when the assignment stops being your problem. You close the tabs. Forget the notation. Become someone who has never heard of question three.

We built an app that interrupts that moment with: “Cool. Why did you do it that way?”

We are students. We understand that this is a betrayal.

Meet GradeProof.

Your homework has a follow-up question

The problem is awkward: AI can help produce a convincing assignment without the person submitting it understanding much of it. A polished PDF tells a professor what was handed in. It doesn’t necessarily tell them what stuck.

An oral exam, or viva, gets closer. Ask someone to explain a choice in their own solution, and there’s something to discuss.

But one teaching assistant cannot spend the rest of the semester interviewing everyone individually. They have other responsibilities, including, occasionally, being a student themselves.

So our app does the interviewing. You submit your written homework as a PDF. An AI reads every page, then asks you questions out loud about the work you submitted.

Yes, we used AI to ask whether you understand homework you might have used AI to write. We are aware of how this sounds.

What happens after you upload?

First, the app checks whether your PDF belongs to the assignment. Then you confirm which pages answer which questions, correcting the suggested matches if needed.

Next comes a short viva: four questions about your own solution, with at most one follow-up each.

These aren’t random textbook questions. The point is to ask why you chose an approach, how it works, or what would change if you tried something else.

You can have each question read aloud. Then you get 20 seconds to think (enough to collect your thoughts, not enough to ask a chatbot), and recording starts.

And the follow-up depends on what you actually said. A vague explanation can lead to a more specific question. “I don’t know” can lead to something simpler about the same part of your work. Follow-ups give you 5 seconds. Good luck.

The examiner has read the submission and is listening to the answer. Terrible news for anyone hoping to survive entirely on “as previously discussed.”

A human still grades the written homework. The viva contributes a separate understanding score, and the instructor chooses how much it counts toward the final grade.

We also needed some follow-up questions

Our first scoring scale started at one.

Even “no understanding shown” had a minimum score above zero. We had accidentally designed participation credit for having absolutely nothing to add.

Generous? Yes. Sensible? We have since discovered zero.

The revised scale also makes full marks reachable without demanding a perfect speech. A short, correct explanation can be enough, and a follow-up that clears up an earlier answer counts in full.

Then there was the voice. Asking it for a “warm, calm tone” made it speak much more slowly.

We wanted a reassuring examiner. We got an examiner with nowhere else to be.

The fix was to stop giving it elaborate delivery advice and let it read the question. Apparently, even our AI needed someone to say, “Please get to the point.”

An interview, not an accusation

This is not a cheating detector. It doesn’t declare that somebody copied, plagiarised, or used AI. It reports what understanding they demonstrated.

That distinction matters. Being nervous is not evidence of cheating. Neither is having an accent or taking a pause.

The assessment instructions explicitly exclude accent, filler words, grammar, and non-native phrasing. Students can also request an instructor-approved accommodation to type their answers instead of speaking.

Scores come with supporting quotes from the conversation. The app checks those quotes against the transcript and flags uncertain or unsupported scores for human review.

Professors can read the conversation, listen to answers, and change scores. Grades stay hidden until the instructor publishes them, and students can request a regrade afterward.

We’re students building this. “The computer said so” would be an especially embarrassing appeals policy.

What comes next?

The unfinished list includes smoother live voice conversations, reusable grading comments and deductions, and better handling of grouped sub-questions.

The app already helps instructors spot concepts that several students struggled to explain. That’s the useful outcome beyond a grade: knowing what needs another pass in class.

Try GradeProof.

Then submit something you understand.