Course experiment scores agents on a replayable contract, not the model

On October 18, 2026, DEV Community published an AI article describing the design of a lab course whose central claim is that a grader should never care which model produced a given agent run. The author sums up the thesis as record once, replay forever, grade offline, and requires every submission to include a replayable record so that scores do not drift when free tiers, rate limits, or model availability change. The article argues that if a rubric silently changes when the free tier changes, what is being graded is the vendor rather than the student. It also notes a failure the author keeps seeing students repeat: around the second week of a bootcamp cohort, students build an agent loop against whatever model happens to be free that morning, get it working, and submit. The report does not name the lab course, the institution running it, or any results from applying the design, and the account stops at the proposal of the approach and the description of the problem.