Start here
AI-Resilient Assessment Kit — three common university assignments, stress-tested against generative AI, then redesigned with explicit AI-use levels, process evidence and new rubrics. Includes an approach comparison, a decision tool and a ready-to-run 90-minute faculty workshop.
Evaluation status: Automated accessibility testing, an agent design review, and agent-simulated stress-test, review-panel and workshop-participant runs (one AI agent, single model) are complete. No real participants, students or educators have been involved. See evaluation/METHODS.md.
The whole kit is set at Lakeside State University — a fictional US public university preparing for ADA Title II digital accessibility compliance. The assignments, rubrics and numbers about enrolment are fictional but realistic; the frameworks and evidence base are real and cited.
What problem this solves
Most AI-era assessment guidance is generic (“use more orals!”, “ban AI”). It doesn’t tell a lecturer which parts of their assignment AI can already do, what to keep, and what to tell students. This kit models the full diagnosis-then-design loop on three realistic cases:
- Stress test the original. Run the brief through AI with three prompt strategies. Find out which rubric criteria AI satisfies easily. (Protocol ready; a smaller agent-simulated run with one Determined output per case is reported on each case page, next to the clearly labelled vulnerability reasoning.)
- Redesign deliberately. Choose an AI-use level for every part of the task (using the AIAS 2.1 levels), move weight onto process and secure checks, and rewrite the rubric to reward what AI can’t do.
- Make it usable. Faculty notes with grading time and scaling options, student-facing briefs in plain language, and a workshop that walks colleagues through their own assignment.
The three cases
| Case | Course | Original assignment | Redesign in one line |
|---|---|---|---|
| A | SOC 101, first-year sociology | 1,500-word take-home argumentative essay | Staged build-up with a source-verification gate and a 5-minute defence |
| B | STAT 201, introductory statistics | Take-home problem set (10 questions) | Open practice lane + weighted secure in-class checks + class-collected data |
| C | PSY 250, research methods | Lab-style research report from provided data | Student-collected data, lab notebook with dated entries, short structured Q&A |
How to use this kit
- If you teach: start with principles, then read the case closest to your own course, then try the decision tool on your own assignment.
- If you run faculty development: take the workshop as-is — guide, slides, canvas, handout and survey are all included.
- If you want the evidence trail: the literature notes and the stress-test protocol (now browsable on this site under Research), and evaluation/METHODS.md.
What’s honest about this build
- Every stress-test score comes from one agent-simulated run: one AI agent wrote one submission per case and scored it itself, not blind, with no second model. The vulnerability analysis in each case is design reasoning, labelled as such, not measured results.
- All reviews so far are AI: an Agent review (a separate AI reviewer in a fresh context) and an agent-simulated review panel (one agent playing three reviewer roles). No educator has reviewed this yet; the review form is ready.
- No detection tools are used or endorsed anywhere in the kit, and no claim of WCAG conformance is made — see the README accessibility section.