An Educator Assistant That Shows Its Work

Most of the work behind a class assessment is not judgment. It is assembly. Writing one more debugging exercise pitched at the right level, pulling twenty questions into a quiz that fits a fifty-minute block, tallying who got what after the room empties out. The judgment, deciding what to assess, what a rough half-right answer deserves, which concept to reteach on Monday, is the part educators actually signed up for. The classroom side of AlgoArena now splits those two kinds of work deliberately. An assistant picks up the assembly. The judgment calls stay with you, and every suggestion the AI makes arrives with something you can check before you act on it.

Ask first, approve second

The classroom dashboard includes an Educator Assistant you can talk to in plain language. Tell it to create a sixty-minute take-home for your data structures class, or to draft a debugging question for second-year students, and it works the way a good TA would. It asks about the context it is missing, student level, topics, time budget, then drafts the questions and proposes the assignment.

The word "proposes" is doing real work in that sentence. When the assistant wants to create or change something, it does not just do it. The proposal shows up as a card in the conversation with a one-line summary of exactly what would change, and nothing happens until you approve it. The assistant is not even allowed to claim credit early; it cannot say "created" or "done" until you have made that true by approving the card. Dismiss the card and the draft evaporates without a trace.

That boundary matters more in a classroom than almost anywhere else. A drafted question costs nothing to regenerate, but an assignment published to forty students is an event you cannot quietly take back. So drafts are cheap and plentiful, and the step that actually touches your class runs through your hands every time.

The assistant also answers questions rather than only building things. Ask how hosting a live quiz differs from student practice, or how your class's recent sessions have gone, and it answers from the product and from your actual data. One thing it will not do is make numbers up. If the analytics are not there, it says so instead of inventing a trend. It also stays on task; ask it for trivia and it will politely steer the conversation back to teaching work.

The score you see is a score we ran

AI-assisted grading has a trust problem in education, and the problem is earned. A model asserting "this looks correct" is an opinion wearing the costume of a fact. Grading in live sessions is built to avoid exactly that.

Coding questions in a live quiz are executed, not eyeballed. When a student submits code, the platform does not take the reported result at face value, not from the student's device and not from a model's summary. For questions with attached test cases, the server runs the submission again itself and uses its own pass counts. For questions judged against heavier test sets, the verdict was already produced on the server and is tied to the exact code that was submitted, so a submission cannot borrow a green result it did not earn. When neither path can verify a result, the score is recorded as unverified rather than promoted to settled fact.

Some question types have no test to run. A whiteboard question asks students to draw, a binary tree, a recursion diagram, whatever the prompt calls for, and rough sketches do not reduce to pass or fail on their own. Here the grader works by comparison. It looks at the drawing against the reference answer, or against the question text when there is no reference, and returns a verdict with a one-sentence reason attached. That reason is the checkable part. You see why it accepted or rejected a sketch, the manual grading buttons stay right where they were, and you can overrule any call, drawing by drawing. If automated grading is ever unavailable, the submission simply waits as pending for your review. The system fails closed, toward the human, never open toward unearned credit.

Four steps to a live quiz

Composing a quiz used to mean a blank page. The builder now walks it as a short stepper, Details, Settings, Questions, Finish, with a progress rail you can click to jump back and adjust anything.

Details is the identity of the thing: what the quiz is called and how it presents to students. Settings is where the pedagogy lives. Topics, difficulty, question count, and timing all sit here, and those choices feed forward, driving both the auto-generator on the next step and the pacing of the live game itself. The Questions step is deliberately plural about sources. You can auto-generate from the same problem bank that powers live competitive play, browse a curated library, have AI generate questions on a custom topic, or type your own, and mix all four in one quiz while reordering and editing as you go. Finish asks the only remaining question: save the quiz to your library, host it live right now, or both.

The formats cover more ground than trivia. Multiple choice sits alongside code challenges run against real tests, debug-and-fix exercises, line-ordering problems where students reassemble scrambled code, and freehand drawing questions.

Watch the answers land, then teach to the misses

Hosting is the payoff. Students join with a short code, the lobby fills, and the session opens with a countdown synchronized from the server, so every screen in the room flips to question one at the same moment. From there, students answer on their own devices under a shared timer while you watch submissions land in real time, and the answer reveal plus a leaderboard between questions keeps the room's energy up.

The part that changes your next lecture comes after the last question. Every session ends with a debrief page. Totals sit up top, how many students participated and the average accuracy across the room, and below them is a participation matrix with a row per student and a column per question, marking each cell as a pass, a partial, or a miss. Click any row to inspect that student's actual answers, question by question. Each question carries its topic, so a column of misses is not a vague sense that the quiz went badly. It is a specific named concept that did not stick, which makes it Monday's lesson plan. When the results need to travel, the whole debrief prints to PDF or exports as CSV for your gradebook.

One rule, three places

Look across these surfaces and the same shape repeats. The assistant drafts, but a change to your class waits on your approval. The grader suggests, but the score rests on checks the platform actually ran, and the calls it could not verify stay yours. The live session collects hundreds of answers, then compresses them into a matrix you can read in a minute and act on tomorrow. The busywork moved to the machine. The judgment did not move at all, and the evidence behind every suggestion is sitting there, waiting for you to check it.