Conversational Assessment

Oral-exam-style assessment at scale. Students talk; a rubric grades.

A student spends a few minutes in conversation with an AI interviewer about a topic. A separate evaluator holds your rubric and the answer key, tracks what the student has demonstrated, and assigns the grade. The interviewer never sees the rubric or the answers, so it cannot be talked, flattered, or argued into revealing them, and the grade reflects only what the student showed unprompted.

How It Works

Two agents, one conversation. The interviewer asks and follows up; the evaluator watches, updates each criterion's status after every message, and steers the interviewer with guidance that never contains the answer. Criteria count as met only when volunteered, not when the interviewer pointed at them.
Your rubric, your scale. Assessments are structured documents: portions, criteria with met / partially-met / not-started definitions, a private answer key, counterarguments to probe with, optional fixed anchor questions, and any grading scale you like.
Authored with your own AI. Connect claude.ai, Claude Code, ChatGPT, Cursor, or any MCP client to your account and write assessments in conversation. The server validates and lints them, then simulates good, weak, and adversarial students against them before any real student sees them.
Delivered where your course already lives. Today: PrairieLearn, with no changes to PrairieLearn itself. Grades land in the gradebook; transcripts are kept for review and appeals.

Get Started

Built and run at the University of Illinois. Sign in with your Illinois account and request access; once approved you can generate tokens, author assessments, and add them to your course.

Method and design notes: the evaluator/interviewer split, scaffolding-aware grading, and the adversarial test harness are described in the authoring guide. Questions: challen@illinois.edu.