In October 2026 I shared a series of LinkedIn‑style posts about redesigning assessment for my Tools in Data Science course at IIT‑Madras, reflecting on AI‑agent adoption, a humorous Scrabble story, and a student success case.
- Learning test – Does solving the question teach the intended concept?
A student achieved 97 % final grade after many failed attempts using AI agents.
Claude generated a Scrabble scenario where the invented word “ZQFMGB” scored exactly 957 points on a double‑word square.
Prototype questions that performed well against the checklist included Prediction markets, Barter assignments, and a Pointless‑style prompt challenge.
- Forward test – Is simple copying of an answer sufficient for all students?
- Solver test – Could a single “solver” be built and shared to give everyone the answer?
- Agent test – Would a large‑language‑model agent (e.g., Codex) solve it when logged in?
- Cheating test – Is cheating more costly than honest effort, or does cheating still yield learning?
- Evaluation test – Is automatic grading easy and robust?