Case A — faculty notes

Workload, scaling, equity and implementation details for the redesign. Everything here is written for the fictional SOC 101 setting at Lakeside State University — a fictional US public university preparing for ADA Title II digital accessibility compliance — with 180 students, one lecturer and two TAs.

Grading time

Component	Original	Redesign	Notes
Per-student marking	15–20 min × 180 (one 1,500-word essay each)	~14–16 min per student × 180 (stages 1–4 below sum to 13–15 min; the sampled defence adds ≈ 1 min per student amortised — see scaling note 1)	Comparable to the original, spread over six weeks; every stage buys evidence the single essay could not
Stage 1 proposal	—	~2 min (checklist)	Rubric-free: complete/revise
Stage 2 source list	—	~4 min (checklist)	Spot-check 1 of 5 annotations against the source
Stage 3 draft + log	—	~1 min (completion)	Log is read properly at stage 4
Stage 4 final essay	(this was the 15–20 min)	~6–8 min	Rubric now carries a verification criterion; no need to hunt citations one by one — the gate already did it
Stage 5 defence	—	~4–5 min live for the sampled fifth of the cohort (≈ 1 min per student amortised; 2 min recorded at 1.5×)	Sampled: see scaling below

Where the grade weight comes from: the five stages total 35% where the single essay was 25%. The extra 10% is re-weighted at adoption from Essay 1 — the course’s other take-home essay, whose weight the course team reduces by 10% on the same reasoning the redesign applies here (its take-home marks are the least secure evidence in the course).

Why it nets out: the totals overlap — roughly 14–16 minutes per student against the original’s 15–20 — so the marking load is comparable, not magically lower. What changes is where the time goes: a little time at five checkpoints instead of all of it inside one artefact, and the verification gate removes the most time-consuming uncertainty (are the sources real?). TAs carry stages 1–3 entirely.

Scaling to 200+ students

1. Sample the defence. Mark the 5-minute defence for a stratified sample (e.g., every 5th student plus any case flagged by the stage 2 checklist or the AI-use log). Non-sampled students still complete it — completion earns the stage mark. Experience from interactive-oral programs (see literature notes) is that the existence of the defence does most of the work; sampling protects the workload.
2. TA-run checklists. Stages 1–3 are checklist-gradable by TAs after a 45-minute calibration session (grade 10 samples together, discuss, re-grade 5).
3. Cohort-level rubric anchors. Before stage 4 marking, the team reads the same three essays together and calibrates criterion scores. Cheap, and it prevents the classic drift between markers.
4. Group-scheduled defence slots. Five-minute slots inside existing tutorial times — no extra contact hours. Recordings accepted for students who can’t attend.
5. Never use detection as a triage shortcut: it adds a false-positive risk the redesign doesn’t need (see principle 6).

Accessibility and accommodations

• The defence always has an equal-weight alternative: live, recorded, or written Q&A (~250 words per question). Students with communication disabilities, anxiety or speaking accommodations choose once at the start of term — no disclosure of diagnosis required, no doctor’s note demanded.
• All stage documents are provided in accessible formats (headings, real lists, no text-in-images) and the LMS pages meet the university’s accessibility baseline; the same checklist used for this kit’s pages works (README accessibility section).
• Extended-time students get the standard extended window on stages 1–4; stage 5’s recorded/written option removes the timing barrier entirely.
• AI-use levels never require paid AI subscriptions: every AI Planning / AI Collaboration stage is satisfiable with the university’s free AI tool or a free tier. Students with no AI access at all complete stages 1 and 3 unaided (the levels permit that) — the log then says so, and that’s a complete log.

Multilingual students

• The stress-test analysis (Case A, criterion 2) and the detection literature (Liang et al. 2023; Weber-Wulff et al. 2023) both argue against detector use; nothing in this redesign depends on one.
• AI Collaboration is explicitly available for language help at stage 3: having AI suggest clearer phrasing of your own sentences, logged, is legitimate and encouraged. What’s assessed at stage 4 is the argument, and the defence verifies the thinking.
• Annotations are graded for content, not idiom: a slightly awkward but accurate annotation scores above a fluent generic one.
• For students still developing academic English, allow the stage 5 written Q&A to count double (defence 10% instead of 5% and essay 10%) on request — same evidence, different medium, no penalty for not speaking.

Common pushback, with answers

• “Five deadlines is more work for me.” The old design was one 15–20-minute mark × 180. The new one is ~14–16 minutes per student × 180 spread over six weeks, with TAs carrying stages 1–3. Marking time is comparable — what changes is what the time buys: five checkpoints of evidence instead of one un-verifiable artefact.
• “Students will just have AI do the draft.” The draft stage is the open lane — AI help there is allowed and logged. What’s assessed is whether the final essay and defence cohere with the student’s own sources and choices. AI-drafted work that the student can’t defend fails the defence, not the honesty rule.
• “The verification gate sounds slow to check.” It’s a checklist: does the annotation reference something specific to the source, is there a page-anchored quote, is the access route named? ~4 minutes including one spot-check.

What we would measure after the first run

• Actual grading minutes per stage (vs. the estimates above) — the estimates are the design targets, not measurements.
• Share of stage 2 checklists that needed a source spot-check second look.
• Defence question completion rates by mode (live / recorded / written) — watching for equity gaps between modes.