The assignment
This problem set finishes and polishes the Redline Labs FY2027 forecast you began in the Session 7 workshop. Your workshop starter.xlsx is the starting state; treat the workshop build as a v0 — nothing in it is certified until you have tied each cell back to a driver and to the model's five checksum cells.
Redline Labs, in one paragraph. Redline Labs sells an AI contract-review copilot to mid-market law firms; it was founded by a former DocuSigner VP of Product. Pricing is hybrid: a $2,500/month platform fee per firm plus usage billed at $2 per document above a 1,000-document monthly allowance. Firms grow their document volume as they mature, 40% of new firms prepay a year in advance at a discount, and the cost of an AI inference falls every quarter. All the exact assumptions live in case_facts.txt and drivers.csv; use those as your source of truth.
AI policy. AI is required on this problem set. Every number in your workbook and memo must be human-certified — you must be able to reproduce and defend each figure without the AI. The closed-book final exam tests the same mechanics (churn, NRR, LTV, driver logic, the cash and deferred-revenue walks) without AI.
The −2 rule. Any number that appears in your memo but cannot be traced to a cell in your model loses 2 points, each.
Worth 40 points: Part A (8) · Part B (18) · Part C (10) · Verification Appendix (4). Both problem sets count (40 + 40).
Group setup
Enter every group member. You need at least 2 and at most 4 names. Put every member's name on all three files.
How this workbench works
- It collects; it does not grade. The workbench records your Part A figures, your five
Checks-tab certifications, your scenario and sensitivity outputs, and your Verification Log. It formats percentages and dollars and flags a few structural things — it never tells you whether a number is correct. - The model stays in Excel. Part B is built in your workbook. Here you certify the five checksum values your model produced.
- Autosave on this device only. Progress saves to your browser. Use Export/Import to move work within your group.
- No data leaves your computer. Nothing is uploaded or tracked. Verify in DevTools: zero network requests on this page.
Part A — Unit-economics warm-up 8 pts
A1 — Churn, logo retention, and NRR 5 pts
Exhibit A is Redline's trailing FY2026 book: the monthly recurring-revenue (MRR) movement and logo movement for the six months July–December 2026. (This is history, provided for the warm-up; it is a separate dataset from the FY2027 forecast you build in Part B.) Using only the figures in Exhibit A:
- For each of the six months, compute the monthly logo churn rate and the monthly logo retention rate.
- For each month, compute the gross MRR churn rate and gross revenue retention (GRR).
- For each month, compute net revenue retention (NRR). Then compute the compounded NRR for the full six-month window. State the NRR formula you used and be explicit about what it includes and excludes.
- In one or two sentences, explain how NRR exceeds 100% even though the book loses roughly 1.5% of its logos every month. Point to the specific line of Exhibit A that drives the result.
Exhibit A — Redline Labs, trailing FY2026 book (Jul–Dec 2026)
Read-only reference. This is the public data for the warm-up.
| Month | Beginning | New | Expansion | Contraction | Churned | Ending |
|---|---|---|---|---|---|---|
| Jul 2026 | 560,000 | 66,000 | 14,000 | (3,500) | (8,400) | 628,100 |
| Aug 2026 | 628,100 | 71,000 | 15,500 | (3,800) | (8,900) | 701,900 |
| Sep 2026 | 701,900 | 80,000 | 17,200 | (4,100) | (11,300) | 783,700 |
| Oct 2026 | 783,700 | 86,000 | 19,000 | (4,400) | (11,700) | 872,600 |
| Nov 2026 | 872,600 | 96,000 | 21,000 | (4,700) | (14,600) | 970,300 |
| Dec 2026 | 970,300 | 106,000 | 23,500 | (5,000) | (14,800) | 1,080,000 |
| Month | Beginning | New | Churned | Ending |
|---|---|---|---|---|
| Jul 2026 | 200 | 24 | 3 | 221 |
| Aug 2026 | 221 | 26 | 3 | 244 |
| Sep 2026 | 244 | 30 | 4 | 270 |
| Oct 2026 | 270 | 32 | 4 | 298 |
| Nov 2026 | 298 | 36 | 5 | 329 |
| Dec 2026 | 329 | 40 | 5 | 364 |
Your monthly grid
Enter your computed rates as percentages. On leaving a cell the field adds a % sign. Amber flag if a month's logo churn + logo retention do not add to 100.
| Month | Logo churn % | Logo retention % | Gross MRR churn % | GRR % | NRR % | Identity check |
|---|
A2 — Spot the AI's errors 3 pts
A group member asked an AI to summarize Exhibit A. It returned the draft below. It contains exactly two mistakes — one is a formula error, the other is a hallucinated benchmark. Find both.
Redline's book expanded steadily through the back half of 2026. Logo churn held near 1.5% per month across the window, so logo retention averaged roughly 98.5%. Gross revenue retention stayed just under 98%, because a handful of accounts contracted and a few churned, but usage expansion from maturing firms more than offset both. December net revenue retention was 111.3%, calculated as ending recurring revenue ($1,080,000) divided by beginning recurring revenue ($970,300) — a very strong result driven by the tenure-based usage curve. For context, according to the OpenView 2026 SaaS Benchmarks Report (Table 9, p. 47), the median net revenue retention for Series A legal-technology companies was precisely 118.6% last year, so Redline is tracking modestly below the peer median but well above the 100% line that separates expanding from shrinking books.
- Identify the two errors. Name which is the formula error and which is the hallucinated benchmark.
- Correct each one. For the formula error, give the corrected figure and the correct formula. For the hallucinated benchmark, state precisely why it cannot stand and what you would do with it (cite the provenance rule from
ground_rules.pdf).
Formula error
Hallucinated benchmark
Part B — The build 18 pts
Complete the FY2027 monthly (Jan–Dec 2027) driver-based model you started in the workshop, in your Excel workbook. Build directly on the starter.xlsx skeleton (tabs Assumptions, Customer & Revenue, Headcount, Income Statement, Cash Walk, Deferred Revenue, Benchmarks, Checks). Every driver comes from case_facts.txt / drivers.csv; honor every rounding convention stated there so your build ties out exactly.
B1. Customer and headcount schedules (6 pts). Build the monthly customer schedule (beginning firms + new logos − churned = ending firms), splitting new firms into monthly-billing and annual-prepay cohorts, and tracking each cohort up the three-step tenure usage curve. Build the monthly headcount schedule — AEs (with the one-month ramp), customer-success managers (at the stated ratio, hired one month early), engineers (with the 60/40 support-vs-core split), and execs. Your Checks tab has five named checksum cells (CHK1–CHK5); CHK1 (December ending active firms) and CHK3 (December total headcount) must tie out exactly. The workshop starter is unverified — audit every row against its driver and against the stated ratios before you trust it.
B2. Income statement and cash walk (6 pts). Build the monthly income statement (platform revenue, usage revenue, COGS — inference, hosting, and CSM payroll — gross profit, S&M, R&D, G&A, operating income) and the monthly cash walk from the $6.6M opening balance (collections in, cash costs out, including commissions paid quarterly). CHK2 (FY2027 total revenue) and CHK5 (December ending cash) must tie out.
B3. Deferred-revenue roll-forward (3 pts). Build the monthly deferred-revenue schedule for the annual-prepay cohort (beginning balance + new prepay billings − revenue recognized = ending balance). CHK4 (December deferred-revenue balance) must tie out. In one sentence, explain the cash-vs-revenue wedge this account creates.
B4. Model hygiene (3 pts). All assumptions live on the Assumptions tab and are referenced by cell — no hardcoded numbers buried inside formulas; every rate and quantity carries a labeled unit; schedules read left-to-right by month with clean beginning/ending links. A grader should be able to change one assumption cell and watch the model re-flow.
Certify the five Checks-tab values
These come from your model's Checks tab. Enter what your workbook produced. The workbench does not know the correct values and never validates them.
Model-hygiene attestation (B4)
Check each box that is true of your workbook. This mirrors the B4 checklist; it is an attestation, not a grade.
Starter audit — corrections you made
The workshop starter.xlsx is a v0 and is unverified. Describe the corrections you made when auditing it against the drivers and stated ratios.
Part C — Scenarios, sensitivity, and the board memo 10 pts
C1 — Three scenarios, through the drivers 4 pts
Build base, upside, and downside scenarios by changing only these four drivers on your Assumptions tab: (i) new-logo quota per AE, (ii) logo churn rate, (iii) documents/firm/month, and (iv) inference cost/document. Scenarios must flow through the drivers so the model re-emits every output; do not plug scenario numbers into the output rows. For each scenario, report end-of-FY2027 cash and months of runway, and give a one-sentence, case-grounded rationale for each driver value (no new facts).
C2 — Sensitivity 2 pts
Produce a one-variable sensitivity table (or tornado) showing how December ending cash responds to each of the four drivers, holding the others at base. Rank the four drivers by their effect on December cash, and confirm the ranking with your own Excel data table — not with the AI's assertion.
Rank 1 = largest effect on December cash. Each rank 1–4 is used once; the workbench flags a duplicate rank in amber. The Δ-December-cash column is optional.
| Driver | Rank (1–4) | Δ December cash ($, optional) |
|---|
C3 — One-page board memo 4 pts
Using one_pager_template.pptx, write a one-page board memo. State the headline, the base case, the single biggest risk, and a pre-agreed re-forecast trigger (the driver and threshold that would make you re-plan). Attach the one-pager exhibit: the active-firm/ARR trajectory chart, the base/upside/downside end-of-year cash table, the top-five assumptions with sensitivity ranks, and your six-KPI dashboard row.
The memo stays in the PDF. The board memo is not entered here — it is the one-page PDF you export from one_pager_template.pptx and upload with your workbook. The −2 rule applies: every number in the memo must exist in a model cell.
Verification Appendix — the log 4 pts
Submit your completed verification log (as a tab in your workbook and rendered in your PDF; this workbench also captures it). The log documents, row by row: the prompt used → the AI's claim → the check you performed → the result (confirmed / corrected / rejected) → the change you made to the model.
Completeness (2 pts). The log covers the Five Manual Checks: extraction fidelity, roll-forward integrity, benchmark provenance, sensitivity replication, and the December column recompute.
Documented catches (2 pts). At least two entries show a real AI error you caught and corrected (a hallucinated benchmark, an invented assumption, a ramp/lag slip, or an internal inconsistency), or affirmative evidence of a check run clean if the AI happened not to err on that line.
Five Manual Checks — coverage
A chip lights when at least one row is tagged to that check and has a prompt, an AI claim, and a check performed. This tracks coverage, not correctness.
The log
| Check tag | Round / role | Prompt used | AI claim | Check performed | Result | Change made to model |
|---|
Review & submit
Check completeness, then generate your submission file. Upload it to Canvas alongside your Excel workbook and your one-page board-memo PDF.