01 — The question
Not which industry makes the hardest AI benchmark. The capstone needs a business decision an undergraduate can understand in ten minutes, model with real evidence, and see the consequences of within a class session.
A multi-unit franchise acquisition, followed by two or three operating decisions. It is the only option where a calibrated economy, a tested engine and computed answer keys already exist.
Choose an industry, acquire real data, build one company, let students decide, simulate the consequences.
Add internal records, contracts and exceptions that ordinary assistants handle badly. Only if a customer needs it.
The program stopped that build. The evidence did not support the scale, and the capstone does not need it.
This briefing is built from the artifacts in this repository, not from earlier summaries. Where a prior claim did not survive inspection, it is corrected on the next screen.
02 — What I investigated
The research did not stop at reading about industries. It downloaded filings and operating records, rebuilt four acquisitions, built one fictional logistics company, and ran two models against it. These are counts of files I opened or outputs I re-ran.
03 — Which industries could work
Each row states the actual acquisition decision, the real evidence behind it, and whether the consequences can be simulated this term. Banking and energy are comparators at the bottom: both have data, neither is a natural first capstone.
| Industry | The student's actual decision | Real evidence we hold | Can we simulate it now? |
|---|---|---|---|
| RestaurantsMulti-unit franchisee | Buy two of four units from a 12-store operator at a defensible price, then improve pricing, staffing and hours over three rounds. | BWW Item 19 median AUV of $3,433,937 across 532 units; 5% royalty and 4% advertising; 7,829 establishment rows joined from Chicago licences and inspections; three scenario packages with code-computed answers. | Best capstone Calibrated 12-site world, 28 passing tests, existing engine and export. |
| LogisticsTruckload carrier | Acquire Cedar Peak, a 2,066-power-unit carrier: decide the true consideration and goodwill, then fund integration inside a 4.55× leverage covenant. | Werner/FirstFleet SPA and exhibits; target FY2025 revenue $631.981m, EBITDA $61.077m; FMCSA 2,520 power units, 409 crashes, 3,192 inspections; 52-document room. | Best alternative No student-facing world. Fleet economics and lease accounting need teaching first. |
| Property managementMultifamily operator | Acquire a manager's portfolio and decide what to fix, what to sell, and whether deposits and deferred maintenance make the price wrong. | HUD names 4,272 management agents; 896 Illinois properties across 273 agents; one real operator with 37 properties and 3,353 units. No rent rolls; address join failed. | Weak evidence Simple economics, but the operating records that make the decision hard are private. |
| ManufacturingPlant or supplier | Buy a plant or supplier: judge capacity, a substitute supplier, inventory and the capital needed to keep the line running. | SEC filings for six registrants; nine unique OSHA sites with 1,028 employees and 2.19m hours; 20 federal awards worth $14.99m. No plant P&L, BOM or supplier records. | Best story, worst evidence The distinctive documents would be almost entirely synthetic. |
| Healthcare servicesSkilled nursing facility | Buy a facility: test staffing, reimbursement, survey exposure and whether the licence and Medicare enrolment transfer on the same date. | Eight CCNs linked across ownership, CHOW, HCRIS, PBJ and deficiencies; 138 buyer links across 17 states, 14,623 beds and $1.711bn net patient revenue. Attribution is unresolved. | Rich data, high cost Regulatory vocabulary would consume the class time available for analysis. |
It has the cleanest identifier chain in the study: Byline's CERT 20624 resolves across 24 quarters, 54 branches, 12 dated mergers and 104 HMDA rows. But the interesting middle layer — loan tapes, core extracts, allowance models — is confidential. Students would spend the session learning regulatory capital rather than making an operating decision.
Duke's 142 plants and 231.18 TWh of 2025 generation are excellent operating data, and 136 of 142 plants match EIA by plant code. But the acquired-plant roster in NRG/LS Power is not public, and no FERC financial join exists. A student cannot acquire a portfolio they cannot name.
Evidence for the counts: Phase 1E cross-industry comparison, Phase 1C/1D reconstructions, Phase 2A transaction cases, Phase 2B financial reconstruction. Percentages are omitted where the denominator is not a population. “No rent rolls” and “no plant P&L” are findings about public data, not failures of effort.
04 — What we found and built
Five tabs: a real filing, a reconstruction, the fictional company, the documents themselves, and what happened when two models read them. Real and generated material is labelled in every tab.
Downloaded from SEC EDGAR, stored and hashed. The exhibit is the target's FY2025 audited financial statements, filed as part of the acquisition. It is the anchor for every number in the fictional company.
| FY2025, $m | Reported | Read |
|---|---|---|
| Revenue | 631.981 | 100.0% |
| Operating profit | 16.457 | 2.60% margin |
| EBITDA | 61.077 | 9.66% margin |
| Net income | 10.492 | 1.66% margin |
| Total assets | 407.714 | 45.7% lease ROU |
| Total liabilities | 261.589 | 64.1% of assets |
| Current ratio | 0.724× | 1.539× ex-leases |
The lease point is the teaching point: a naive liquidity read says distressed; the underlying working-capital position is +$30.693m.
Public filing, retrieved 4 November 2026. The acquisition closed 27 January 2026. The source file is retained with its SHA-256 in the source manifest.
Local source: fenrix-corporate-data/sources/phase2a/logistics/werner_ex99-1_target_fy2025.htm (35,696 bytes). The reconstructions below use the merger agreement, both 10-Qs, the target interim statements and the FMCSA files.
| Werner / FirstFleet bridge | $m |
|---|---|
| Announced equity price, incl. maximum earnout | 245.000 |
| Add real estate purchased separately | 37.800 |
| Gross transaction value | 282.800 |
| Less finance lease liabilities assumed | −57.100 |
| Net before deferred revision | 220.623 |
| Less deferred-cash reduction | −5.868 |
| GAAP consideration transferred | 214.755 |
One disclosed residual: the $5.912m SPA escrow and the $5.868m deferred-cash reduction differ by $0.044m. It is reported, not smoothed.
Path: reports/industry_selection/phase2b/FIRSTFLEET_FINANCIAL_RECONSTRUCTION.md
Revenue scaled to 0.82× the audited FirstFleet base. Every cost ratio is held or explicitly shifted; the shifts are labelled DESIGN, not randomized. The acquisition bridge and the P&L are internally consistent to the cent.
| Cedar Peak, $m | FY2025 | Basis |
|---|---|---|
| Revenue | 518.225 | ANCHORED × 0.82 |
| Operating profit | 17.485 | derived |
| EBITDA | 53.815 | derived |
| Headline equity price | 230.000 | DESIGN |
| GAAP consideration | 203.400 | bridge |
| Goodwill | 18.152 | residual |
| Post-close stub | 155 days | 27 Jan – 30 Jun |
| Recognised H1 revenue | 226.670 | ledger less in-transit |
01_corporate 8 memo, SPA, escrow, board 02_financial 12 statements, TB, AR/AP, debt 03_operational 10 fleet, safety, terminal 04_contracts 8 MSA, amendments, invoice 05_integration 6 budget, status, org 06_tax 5 nexus, DTL, revenue policy 07_correspondence 3 email threads
Formats: 18 CSV, 18 DOCX, 8 XLSX, 4 PDF, 3 EML, 1 JSON. 47/47 QC checks pass; 107/107 hidden-answer self-checks pass. The manifest marks every file synthetic: true.
These are the real generated documents, embedded in full — not mockups. Each is a rendered copy of the DOCX or PDF that the pipeline produced and the QC suite hashed.
These are illustrative documents for a fictional company. They are not records of any real carrier, employee, supplier or customer. The CFO letter and statements are calibrated to the real FirstFleet ratios, but the entity and every counterparty are invented.
| Run | Acceptable | Raw | Adjusted |
|---|---|---|---|
| terra B | 6 / 8 | 94/107 | 99/107 |
| terra C | 7 / 8 | 99/107 | 104/107 |
| sol B | 6 / 8 | 98/107 | 103/107 |
| sol C | 7 / 8 | 101/107 | 106/107 |
Condition A (no documents) abstained on all 16 task-runs and fabricated no numbers. Both models were OpenAI variants; there was one run per cell.
The useful result is not the score. It is that the harness can tell a model mistake from a bad question — which the earlier summary did not distinguish.
| Defect | Count |
|---|---|
| Wrong answer key | 1 |
| Ambiguous denominator | 1 |
| Percent / fraction mismatch | 1 |
| Missing document in task set | 1 |
| Sign-convention ambiguity | 1 |
| Single-document shortcut | 3 |
terra B reported PPA assets of $436.125m. The key is $430.257m. The difference is exactly the $5.868m deferred-cash reduction — the model added it back into assets instead of deducting it from consideration. Cash paid and goodwill tied, so this is a reconciliation error, not a retrieval failure.
The key says replacing $13.2m of rent with a finance lease lifts EBITDA by $12.0m. The correct uplift is $13.2m. terra returned 13.2 and was marked wrong. The instrument, not the model, failed. Both are now recorded separately.
This is why the recommendation is not “logistics is validated.” The pilot is good enough to show the machinery works and to find defects. It is not good enough to prove enterprise reasoning: 6 of 8 tasks were cleanly solvable, three had one-document shortcuts, and only one model family was tested.
05 — What the student would do
Each is a 60–120 minute analytical exercise with a defensible answer. Figures marked “illustrative” are design choices; figures with a source are anchored in real data.
The three concepts deliberately differ in data readiness. Restaurants have real anchors for revenue, cost and capital, and a missing franchisee four-wall margin that must be treated as a parameter. Logistics has the strongest real transaction and the weakest student familiarity. Property has the simplest economics and the weakest operating evidence.
06 — Recommendation
The first four criteria carry 90% of the weight. Commercial extension is a tie-breaker, not the reason to choose an industry.
Not because the AI problem is hardest. Because it is the only option that can be a playable, gradeable business decision this term without inventing the economics from scratch.
Choose it if the team values a faithful real acquisition over classroom speed. The Werner/FirstFleet case is reconstructed to the dollar, and the 52-document room exists.
It can be a separate track from the capstone. The document work that makes it hard to teach — amendments, entity maps, price bridges — is the work an enterprise assistant is actually tested on.
Do not rebuild Springfield around another vertical. Do not generate 650 documents. The evidence does not support either decision.
Weights: accessibility 30, data readiness 25, decision quality 20, simulation feasibility 15, commercial extension 10. The ratings are qualitative and evidence-linked; no 91-versus-90 score is asserted. The 91.6% shortcut figure is measured in research/springfield-econ/reports/VALIDATION.md.
07 — Commercial extension and team decision
What we proved is narrow: a harness can index a mixed-format room, models can find controlling documents, and the scoring can separate a model error from a bad key. What we have not proved is that anyone will pay for it.
There is no customer validation, no willingness-to-pay evidence and no proof of training value. The benchmark is currently too easy and too cooperative to support an enterprise-reasoning claim. Rights are unresolved: WRDS prohibits general generative-AI use, and public SEC access is not a redistribution or training licence.
What would count as evidence: a named design partner, a paid pilot with a defined workflow, a rights determination from counsel, and a task set that makes a competent model fail for the right reason.
Supporting files — README, EVIDENCE, CAPSTONE_OPTIONS, COMMERCIAL_EXTENSION, ARTIFACT_GALLERY, SLACK_DRAFT — are in the GitHub package. This copy is self-contained: the four document previews above are embedded and need no external files.
Fenrix capstone industry selection · 5 November 2026 · built from the local research workspace. Paths beginning sources/, reports/, releases/, normalized/ or scenarios/ are relative to fenrix-corporate-data/; research/ and docs/ are relative to the project root.