Case B — faculty notes

Cost, scaling and equity notes for the redesign. Setting is the fictional STAT 201 at Lakeside State University — a fictional US public university preparing for ADA Title II digital accessibility compliance — 250 students, one coordinator, four TAs.

Grading time

Component	Original	Redesign	Notes
Weekly sets ×5	10–12 min × 250 × 5 weeks (≈ 210–250 h per term across TAs)	~1.5 min × 250 × 5 (≈ 31 h) — completion + log check	The answer key stops mattering; the log is 5 lines
Secure checks ×2	(nothing equivalent before the midterm)	~4 min × 250 × 2 (≈ 33 h)	Four questions each; Q4 is 1 mark; marks fast
Net	≈ 210–250 h	≈ 64 h of marking + calibration (≈ 4 h) ≈ 68 h	Roughly 30% of the original marking load (≈ 68 h vs ≈ 210–250 h), with far more assurance

Where the grade weight comes from: the original five sets were worth 10% each (≈ 50% of the course grade — see the original brief). The redesign cuts the practice lane to 10% total and moves 20% of the freed weight into the two secure checks; the sets-and-checks block drops from ≈ 50% to 30% of the course grade, and the remaining 20% is re-allocated by the course team at adoption (at Lakeside State, to the final examination). The midterm (20%) is unchanged.

The design trades marking answers for marking logs and short checks. If logs feel too light, the alternative is sampling a deep read of 10% of sets per week (25 scripts × 10–12 min ≈ 4–5 h per week, ≈ 20–25 h across the term).

Scaling to 200+ students

The secure checks run inside existing tutorial slots — no invigilation hiring, no room booking. Sections of 30 take 20 minutes with the TA present; nothing else fits their slot that week.
Log grading is binary (complete/partial) — a checklist, not an essay mark. New TAs calibrate on 15 sample logs in one 45-minute session.
The class dataset is collected once (week 2, LMS form) and cleaned by the coordinator; every section uses the same file. No per-section variation to manage.
“Find and fix” items are pre-written (from last year’s anonymised errors); adding one to a set costs nothing.
If TAs are scarce, the checks can be marked on a 20-point scheme by Q4 auto-marking (LMS quiz format for Q1 and Q4, hand-marked for Q2–Q3) — about 2 min/script.

Accessibility and accommodations

Secure checks: any student with extended-time or rest-break accommodations takes the check with the TA’s standard arrangement; because the check is only 20 minutes, doubling time still fits a tutorial slot.
Missed check: a make-up slot in the same week; a second miss converts to a 15-minute oral equivalent (or written Q&A) with the coordinator — accommodations don’t require disclosure of reasons.
Open lane: logs can be voice notes, bullet points, or table rows — format doesn’t matter, honesty does. Screen-reader users and students with RSI can use AI more, not less: the levels never require handwriting, and the tutorial checks have a typed equal-weight alternative per the university’s accommodation process.
The class survey is optional to answer (it informs the dataset, not your grade); students who opt out use the anonymised dataset like everyone else. No sensitive questions; the collection form is reviewed by the teaching team against the university’s minimal-data guidance.

Multilingual students

Interpretation-in-context sentences are graded on meaning, not idiom — a plain, slightly awkward sentence that gets the numbers and limits right scores full marks.
The open lane explicitly legitimises AI language help (“ask AI to explain the question in simpler words” is a logged, allowed use) — which is exactly the support multilingual students have been using silently under the old ban.
Detector use stays off the table: the evidence on false positives for non-native writers (Liang et al. 2023; Weber-Wulff et al. 2023) is the reason the redesign needs no detector.
The formula sheet may include the student’s own brief wording of each formula in any language — transcription into English is not being tested.

Common pushback, with answers

“Full AI on problem sets is surrender.” The sets stop certifying anything; the checks certify. The alternative — banning AI from unenforceable take-home sets — certifies nothing and manufactures suspicion.
“Two 20-minute checks can’t measure everything.” Correct. They measure the four highest-priority skills, weighted honestly (20% > the practice lane’s 10%), and the midterm remains.
“Students will split logs — write the log after copying answers.” Some will; that’s why the checks exist and why the log grades judgment (catching an AI error) rather than compliance. The redesign doesn’t claim to make everyone honest; it makes honesty the cheap, safe strategy and makes certainty possible on the secure lane.

What we would measure after the first term

Check-score correlation with midterm score (validity signal for the checks).
Log-quality distribution and its relationship to check performance (does engaged AI use predict secure-lane scores?).
TA hours actually spent vs. the estimates above.