Evidence vs. Opinion I
The Captive Finance Trap
Same workbook. Same valuation task. Very different levels of discipline.
A global car maker that also lends money to its customers. One workbook, one question, and three very different answers.
The difference was not confidence. The difference was evidence.
Claude Run 1
$170.88/share
Claude Run 2
$185.48/share
Sharpe AI
$125.68/share
Core issue
Valuation bridge discipline
Biggest lesson
The bridge was the model
Looks professional
Statements tie. Checks pass. Formatting is clean.
Logic under review
The model looked professional. That was the problem.
Chapter 01
Three answers, one workbook
Two AI tools were handed the same workbook and asked the same question: what is this company worth?
Here is what came back.
| Run | Valuation frame | WACC | Terminal method | Value / share |
|---|---|---|---|---|
| Claude Run 1 | Industrial DCF + finance arm at 7.0x earnings | 9.51% | Perpetuity growth: 1.5% | $170.88 |
| Claude Run 2 | Industrial DCF + finance arm at 0.8x book | 10.67% | Perpetuity growth: 2.0% | $185.48 |
| Sharpe AI | Industrial DCF + finance stream at derived book equity | 5.71% | Exit multiple: 8.32x EV/EBITDA | $125.68 |
Same file. Same instruction. Three defensible-looking conclusions.
A clean spreadsheet is not the same thing as a clean valuation. Sometimes it is just a very well-dressed problem.
The result is not a story of "higher number versus lower number." The better question is:
Which valuation could survive a serious review?
Chapter 02
The trap: a company inside a company
A car maker is one thing. A car maker with a finance arm is another.
One group holds two different businesses: industrial operations, and a lending operation with its own receivables, debt, and equity. Some balance-sheet items belong to one, some to the other. Some are already netted. Some should be added back, some subtracted, some excluded.
And if that logic is not explicit, the model can quietly count the same economics twice.
Group company
One workbook. Three kinds of economics.
Industrial operations
Industrial DCF
Enterprise value
Financial services arm
Derived finance stream value
Separate bridge treatment
Cash, debt, minority interest, investments
Equity bridge
Value per share
That is why this case was never just about building a DCF. It was about controlling the model boundary.
Chapter 03
Two analysts in one tool
Claude deserves credit. It did not build a naive consolidated DCF. It saw the captive-finance issue and moved toward a sum-of-the-parts valuation. That is a good finance instinct.
But run it twice, and the process shows its looseness:
What changed across Claude runs: same workbook, same task
| Choice | Claude Run 1 | Claude Run 2 | Why it matters |
|---|---|---|---|
| Value / share | $170.88 | $185.48 | Same workbook, different conclusion |
| WACC | 9.51% | 10.67% | Discount rate changed materially |
| Finance-arm value | 7.0x earnings | 0.8x book | Different method for the same stream |
| Terminal growth | 1.5% | 2.0% | A major value driver, moved between runs |
| Bridge treatment | Receivables / debt shortcut | 70% industrial cash assumption | Bridge choices drove major value differences |
Not one of these choices is indefensible on its own. That is exactly the problem: a smart first draft is still not a controlled modelling workflow.
Claude produced two different analysts on the same file: impressive productivity, slightly awkward for sign-off.
Chapter 04
Process over confidence
Sharpe AI did not win this case by typing a more conservative number. It behaved differently because every major choice had to trace to workbook evidence, pass a test, or surface as a review item.
Read the workbook
Extract the evidence
Evidence-linkedTest forecast drivers
Tested against historyBuild the DCF
Scan the bridge
Bridge checkedVerify the output
Human review flagsThe forecasts were tested, not narrated. Before using a method, the engine asks:
Would this method have worked on the company's own past data?
In this run, 19 drivers were committed. Auto revenue passed with a backtest score of 0.368, far better than the naïve baseline. Motorcycle shipments did not pass cleanly, so the line was not dressed up: it was carried as Review
The system does not need every line to look perfect. It needs weak lines to stop pretending.
How a forecast method is tested
Forecast test = method error ÷ baseline error
- Below 1.0: the method beat the simple baseline.
- Above 1.0: it performed worse than the baseline.
This is not meant to replace human judgement. It is meant to stop unsupported judgement from sneaking into the model dressed as evidence.
Evidence testing in this run
| Line item | Evidence path | Result | Treatment |
|---|---|---|---|
| Auto revenue | Ratio to external operating metric | 0.368 | Accepted |
| Motorcycle shipments | Operating driver screen | Failed the bar | Review |
| Finance stream value | Assets less liabilities from workbook | $19,313m | Accepted |
| Terminal value | Workbook-linked EV/EBITDA level | 8.32x | Cross-checked |
Chapter 05
The bridge was the model
Most valuation debates start with the forecast. In this case, the bridge was more dangerous.
If a finance receivable reduces group debt and the finance arm is added back as a separate value, the same economic asset works twice. If a cash allocation is typed without evidence, a single blue cell moves the value per share dramatically.
Here is Sharpe AI's bridge, with every claim included, excluded, or explained:
Enterprise value
$152,598m
Finance stream value
+$19,313m
Cash
+$25,033m
Investments
+$4,681m
Debt
($121,921m)
Minority interest
($2,594m)
Equity value
$77,110m
÷ shares outstanding
$125.68
Amber markers: bridge items where double-counting or unevidenced allocations can quietly move value.
In complex valuations, the bridge is not admin. The bridge is the model.
Chapter 06
The verdict
This case is more credible if we are honest about both sides.
What went right
- Claude correctly noticed that a simple consolidated DCF was not enough.
- All runs moved toward a sum-of-the-parts style approach.
- Sharpe AI tied major values to workbook evidence, tests, or review flags.
- Sharpe AI treated the bridge as a controlled workflow, not a loose adjustment block.
What stayed risky
- Claude run-to-run drift: the same workbook produced two materially different answers.
- Bridge assumptions: cash, debt, receivables, and finance-arm treatment moved value materially.
- Captive finance businesses are easy to mis-handle if the model boundary is unclear.
- Sharpe AI record issue: one save was refused because a summary number contradicted the compiled sheet. That is not hidden. The check worked.
The last point matters. Sharpe AI is not claiming the model can never make a mistake. The claim is different:
If something breaks, the system should catch it, flag it, or refuse to silently publish it.
Because a serious reviewer never asks only "what is the company worth?" They ask:
- Where did the number come from?
- What evidence supports it?
- What was tested?
- What was a judgement call?
- What still needs human review?
A model can pass the eye test and still fail the review test. A financial model is not finished when Excel returns a value. It is finished when a serious reviewer can attack the logic and the model is still standing.
That is what a financial model control layer is for.
Appendix
The full workings
The complete numbers behind chapter 05, for readers doing diligence.
For the technical readerSharpe AI valuation chain: every step to $125.68
Sharpe AI valuation chain
| Step | Value | Comment |
|---|---|---|
| PV of explicit FCFF | $21,210m | Forecast period cash flow value |
| Terminal value | $173,428m | Exit multiple method (undiscounted) |
| Enterprise value | $152,598m | DCF value before claims bridge |
| Add finance stream value | $19,313m | Derived from assets less liabilities |
| Add cash | $25,033m | Workbook-linked bridge item |
| Add investments | $4,681m | Included in bridge |
| Less debt | ($121,921m) | Frame-tied claim |
| Less minority interest | ($2,594m) | Claim on equity value |
| Equity value | $77,110m | Final equity bridge output |
| Value / share | $125.68 | Sheet-evaluated value |
For the technical readerWhere the bridge could go wrong: the risk map
Bridge risk detail
| Issue | Why it matters | What Sharpe AI did |
|---|---|---|
| Finance-arm value | Can double-count assets or debt | Derived separately and tracked explicitly |
| Debt treatment | Major driver of equity value | Forced through bridge logic and claim checks |
| Cash allocation | Can overstate industrial equity value | Treated as a judgement that must be evidenced |
| Share count | Directly affects per-share output | Checked against the final output basis |
| Missed claims | Can silently distort equity value | Material claims included, excluded, or explained |
Want to see what Sharpe AI would do with your model?
Bring a DCF, a messy workbook, a first draft from AI, or a model your team needs to defend.
For product demonstration only. Not investment advice.