About the unverified marks on this pageFigures tagged unverified are PACE-local estimates that could not be confirmed against any published source, so they are labelled rather than quietly presented as fact. Every other factual claim on this page — statutes, standards codes, gene biology, primary-source citations — was independently checked; the full claim audit, including what came back wrong, lists every claim with its source.
Try it Tuesday · 90 minutes
Run this next week
90 minutes, one decision that has two sides to it — a lease renewal, an acquisition price, a covenant a lender wants and a borrower does not. Minutes 0–5: split the room. Half model the decision as the party who wants it to happen; half model it as the party who has to be talked into it. Neither half sees the other’s work. Minutes 5–35: each side builds its case with the prompt or the embedded break-even demo and commits, in writing, to the one number it will defend. Minutes 35–80: both numbers go up side by side, against a single question — what accounts for the distance between them? Work the gap down input by input until it closes or the room agrees it cannot. Minutes 80–90: what gets handed in is the reconciliation — the short list of inputs that produced the spread, which of them a filing or a quote could settle, and the one that is a view about the future no model will ever decide. Finance students should leave having watched two honest analysts with the same tool arrive at different numbers, and knowing that saying precisely why is the professional act.
The full assignment
Make it Laser cut · engrave
Physical decision cards or a data object for a simulation: a term structure or a risk distribution rendered as something you can hold and compare side by side. Useful precisely because it slows the reading down.
Build it AI chat · no code
A working interactive model for a real decision, built live. The assignment is not the model. It is the sensitivity analysis and the written defense of every assumption the model needed to make.
Try the demo · runs offline
Finance demo
Interactive demo — “break-even”. It cannot be printed; run it in a browser at https://el3vate.vercel.app/demos/break-even/index.html
This demo runs entirely offline against canned data — no keys, no network, safe on conference wifi.
5 weeks, realistically
The 5-week version
- Wk 1Pick a real decision with public filings. Students list every assumption required, before building anything.
- Wk 2Build the interactive model with AI in class. Break it deliberately: what input makes the conclusion flip? By the end of the week the term structure or risk distribution the cards will render is fixed — lay the cards out from it.
- Wk 3Cut file for the physical comparison cards due the first day of this week. The sensitivity write-up runs while they are cut and engraved.
- Wk 4Teams rehearse the defense and mark which assumption they expect to be attacked. The cards come back this week.
- Wk 5Defense session. Another team’s job is to find the assumption that carries the whole result.
Starter prompt & what the model actually did
The prompt, run for real
Build a single self-contained HTML page with an interactive model for this decision: [describe the decision]. Include sliders for every input that matters. Show the output updating live, and display the break-even point prominently. Then list every assumption you had to make to build it, and tell me which single assumption the conclusion is most sensitive to.
Here's a self-contained page. Save as `chinatown-lease.html` and open in any browser.
```html
<!doctype html><meta charset="utf-8"><title>Chinatown Lease Model</title>
<style>
body{font:15px/1.5 system-ui;max-width:640px;margin:2rem auto;padding:0 1rem}
label{display:block;margin:.6rem 0 .1rem;font-weight:600}
output{float:right;font-weight:400;color:#555}input[type=range]{width:100%}
#out{background:#f4f4f5;border-radius:8px;padding:1rem;margin:1.2rem 0}
.big{font-size:1.4rem;font-weight:700}.g{color:#128a3a}.r{color:#c0392b}
</style>
<h1>Should I sign the Chinatown lease?</h1>
<div id="c"></div>
<div id="out"></div>
<script>
const F=[
["rent","Monthly rent ($)",2000,12000,6500,100],
["ticket","Average ticket ($)",8,40,16,.5],
["covers","Daily covers",30,400,120,5],
["food","Food cost (%)",20,50,32,1],
["wages","Monthly staff wages ($)",0,25000,9000,250]];
const S={},c=document.getElementById("c");
F.forEach(([k,lbl,mn,mx,v,st])=>{
c.insertAdjacentHTML("beforeend",
`<label>${lbl}<output id="o_${k}"></output></label>
<input type=range id=${k} min=${mn} max=${mx} value=${v} step=${st}>`);
S[k]=document.getElementById(k);
S[k].oninput=calc;});
const DAYS=26; // open days/month
function calc(){
const v={};F.forEach(([k])=>{v[k]=+S[k].value;o_(k,S[k].value);});
const rev=v.ticket*v.covers*DAYS;
const gross=v.ticket*(1-v.food/100); // margin per cover
const fixed=v.rent+v.wages;
const net=rev*(1-v.food/100)-fixed;
const beCovers=fixed/gross/DAYS; // covers/day to break even
const cls=net>=0?"g":"r";
document.getElementById("out").innerHTML=
`Monthly revenue: $${rev.toLocaleString()}<br>
Monthly net profit: <span class="big ${cls}">$${Math.round(net).toLocaleString()}</span><br><br>
<span class="big">Break-even: ${beCovers.toFixed(0)} covers/day</span>
<br>(you're modeling ${v.covers})`;
}
function o_(k,val){document.getElementById("o_"+k).textContent=
(k=="food"?val+"%":k=="ticket"?"$"+val:(+val).toLocaleString());}
calc();
</script>
```
**Assumptions I had to make:**
1. **26 operating days/month** (closed ~1 day/week).
2. **Break-even is contribution-based**: only rent + wages are fixed; food cost scales with revenue. All other overhead (utilities, insurance, POS, permits, waste, marketing, loan/build-out) is set to **zero**.
3. **Wages are fixed**, not scaling with covers.
4. **No ramp-up** — projected covers apply from month one, no seasonality.
5. **The lease itself (3-year commitment, deposit, escalators) carries no modeled cost** beyond monthly rent, and the truck's existing revenue isn't netted out as opportunity cost.
6. Average ticket and food-cost % are constant across all covers.
**Most sensitive assumption: daily covers.** Because it's the only lever that multiplies *both* revenue and the food-cost deduction, net profit moves faster with covers than with any single dollar input — and covers is also the least certain (a truck's foot-traffic doesn't transfer 1:1 to a fixed address). A 20% miss on covers swings the outcome more than a 20% move in rent or ticket. Pressure-test that number before signing.
- Self-listed six assumptions, including 'all other overhead... set to zero'The model surfaced its own load-bearing simplifications — the assumption inventory the assignment demands, done unprompted.
- Named the most sensitive assumption: daily coversCorrectly reasoned that covers multiplies both revenue and cost deduction and is least certain — the sensitivity insight the exercise grades.
- Zeroed overhead silently in the formula, then admitted itThe break-even is contribution-based and omits utilities, insurance, permits, and build-out; a student trusting the headline number without reading the note would be badly misled.
- Ignored the lease's own terms (deposit, escalators, three-year commitment)It flagged that the multi-year commitment 'carries no modeled cost beyond monthly rent' — a real omission for a decision that is fundamentally about the commitment.
Assessment
Rubric
| Criterion | Weight | What it assesses |
|---|---|---|
| Fidelity to the assigned side | 25% | The team argued the position it was dealt in good faith, choosing inputs a real party on that side would actually put its name to, rather than drifting toward the middle to look even-handed. |
| Decomposition of the spread | 35% | The distance between the two numbers is traced to named inputs and quantified — this much of the gap comes from the rent line, this much from the ramp — instead of being written off as a difference in outlook. |
| Evidence versus judgment | 25% | Separates the inputs a filing, a quote or a comparable could settle from the ones that rest on a view about the future, and says plainly which is which. |
| Refusal to split the difference | 15% | Moves off a number when the other side's evidence beats it and holds when it does not; an average of the two opening positions earns nothing. |
What this replaces
Swapping it into a real course
Replaces
The valuation or capital-budgeting spreadsheet submitted as the answer.
What is lost
The hand-built modeling mechanics and formula fluency.
What is gained
With AI building the model in minutes, class time moves to the actual professional skill — defending assumptions and finding the one the conclusion hinges on — instead of grading spreadsheet arithmetic a tool now does.
Where AI is bad at this
The failure your students should catch
A model builds a financial model that runs cleanly and hides its load-bearing assumptions — silently setting overhead to zero, assuming no ramp-up, holding margins constant, and presenting a confident break-even without surfacing that one uncertain input drives the result. Its output looks decision-ready while burying exactly the assumptions a sensitivity analysis is supposed to expose.
Budget & logistics
What it costs to run
- Instructor prep1.5 hours
- Class time90 minutes
- Per-student cost$0 for the Tuesday version; roughly $8–15 per team in laser stock unverified for engraved decision and comparison cards in the five-week build.
- Fabrication file dueFirst day of Week 3, in a five-week plan — the comparison cards render the term structure or risk distribution the week-2 model produces, so they are laid out at the close of week 2 and the cut file goes in the moment week 3 opens. PACE quotes 7–10 business days (up to 14 calendar days), which clears with the week-5 defense exactly 14 days out.
- Calendar dependencyModel-building needs no lead time. This plan is five weeks, not four, and the deadline sits at the start of week 3 rather than the end of it — both deliberately. The cards render numbers that do not exist until the week-2 model does, which rules out the start-of-week-2 deadline every other fabrication assignment here uses; and an end-of-week-3 file would leave only 8 calendar days before a week-5 defense, which the up-to-14-calendar-day cut turnaround does not fit. Start of week 3 to a week-5 session is 14 days, which does. A four-week version would hand out cards after the defense they exist to inform.
Three sizes
Scale it to the time you have
One session
The 90-minute two-sided session producing a written reconciliation of the spread between the room's two numbers; no fabrication.
5 weeks
The seeded plan runs five weeks, not four, because the comparison cards render numbers that do not exist until the model does. Week 1, pick a real decision with public filings and list every assumption it requires. Week 2, build the model with AI and break it deliberately; by Friday the term structure or risk distribution is fixed, so the cards get laid out from it. Week 3, cut file in on day one, with the sensitivity write-up running while the cards are engraved. Week 4, teams rehearse the defense and mark the assumption they expect to be attacked; the cards return. Week 5, the defense session. The extra week is the cut turnaround, not padding.
One semester
A full decision case with public filings — assumption inventory, an AI-built model stress-tested repeatedly, fabricated comparison cards, and a defense session where another team must find the assumption carrying the result.
Reserved · live build
This space is intentionally empty. During the Day 8 session it will be filled in live — fill the liveBuild field in content/finance.json and rebuild.
Tell us what happened
Run it, then say how it went
If you try this — the 90-minute version, the prompt, any part of it — send back what you tried and what happened, especially anywhere the model was confidently wrong. That is the material the next session is built from.
Opens a draft in your mail client, already addressed and titled. No form, no account, nothing to sign up for.