The Price Got in the Way

This company was not trading normally.

The share price was pinned by a pending takeover. That matters because a takeover price is not just a view on standalone cash flows. It can include control value, strategic value, scarcity value, and negotiation dynamics.

A DCF built from the company's own operating history should not automatically bend toward that price.

In fact, if the fundamentals are weak and the price is high, the model should be allowed to say:

That is what made this workbook valuable as a test.

It was not just a valuation exercise.

It was an anchoring experiment.

The business had just come through a trough year. Revenue was down 30% to $110.5m. EBIT margin had compressed to 5.6% from a much stronger historical level. Bookings had declined from 192m to 115m over four years. Yet the stock was sitting at $55.08/share, roughly 58x trough EBIT, because a control transaction was pending.

The share price walked into the room wearing a very expensive suit.

The model's job was not to be impressed.

The most important part of this case is that the numbers should not collapse into one answer.

There were three different layers:

A weak model tries to make these numbers collapse into one answer.

A serious model keeps them separate.

The base case should show what the evidence supports.

The thesis case should show what the analyst believes could happen.

The market price should be compared, not smuggled into the assumptions.

Claude did not simply copy the market price.

That would be the lazy criticism, and it would be wrong.

Both Claude runs actually concluded that the market price was above standalone value. That is good.

But Claude's base case still carried a recovery thesis: revenue rebounding after a bad year, margins improving toward prior levels, and cash flows rising materially over the forecast period.

The problem was not that the thesis was absurd.

The problem was that the thesis could not hold still.

Same workbook. Same task. Two runs.

One answer: $36.66/share. Another answer: $27.06/share.

That is a 26% gap inside the same tool.

The WACC change is the cleanest exhibit.

A size premium is not a ridiculous choice. For a small-cap company, it can be perfectly defensible.

That is exactly the problem.

A defensible choice appeared in one run and disappeared in another. The user did not choose a policy. The model chose a mood.

Claude's output was not useless. In fact, this was one of its strongest performances in the series. It recognized the takeover distortion. It built a plausible recovery case. It produced a useful first-pass valuation. It even said the market price likely reflected a control situation rather than standalone cash flows.

That is good analyst instinct.

But a first-pass valuation is not a controlled model.

A recovery thesis needs to be labeled as a thesis.

It cannot quietly become the base case.

Sharpe AI's base case did not try to explain the market price.

That was the point.

It read the workbook, tested the company's own history, and built the base case from what the evidence could support. The result was much lower: $15.03/share.

This does not mean Sharpe AI "knew" the company was only worth $15.

It means the system refused to put a recovery story into the base case unless the workbook evidence supported it.

The recovery story belongs somewhere else: in a named scenario.

That is the product lesson.

Sharpe AI should not eliminate human judgment. It should force judgment to wear a name tag.

This is why the output is useful to a finance team.

It does not merely say "low" or "high."

It separates evidence, thesis, and market price.

Anchoring does not mean the analyst is foolish.

It means the analyst sees a number and then builds around it.

Everyone in finance does this. That is why process matters.

If you know the company is trading at $55.08/share, and you also know your standalone DCF is coming out far below that, your brain starts asking helpful-sounding questions.

The careless version of the claim would be: Claude copied the market price.

That is false.

Claude did not back-solve to $55.08/share. Both runs concluded the market was expensive relative to standalone value.

The real issue is subtler and more universal:

Claude wrote a recovery thesis while the price was already in the frame.

That is how human analysts behave too. The point is not that Claude is unusually biased. The point is that a finance workflow needs architecture that protects the base case from the anchor.

Sharpe AI's driver-fitting machinery did not read the market capitalization while building the operating forecast. The price came later, as a comparison.

That is not superior discipline.

It is architecture.

This was one of the strongest parts of the Sharpe AI run.

The system first tried to build a fundamental terminal value. It did not simply jump to a convenient terminal growth assumption.

But the stable-state checks rejected the fundamental build.

The reinvestment rate was −256.9%. Terminal capex / D&A was 0.29. In plain English: the terminal year did not look like a sustainable steady-state business. So Sharpe AI refused to let that method quietly flow into the model.

Instead, it used the sanctioned fallback: a 2.0% terminal growth rate anchored inside the permitted macro / risk-free range. Crucially, it printed the failed diagnostics next to the number actually used.

That is the product.

Not: AI picked 2%.

But: AI tried the stronger method, rejected it, explained why, used a fallback, and left the evidence on the page.

This case should not pretend Sharpe AI was perfect and Claude was useless.

That would be less credible.

Claude's recovery thesis was not worthless. In fact, Claude's bull case was useful because it captured the upside regime that a simple evidence-led base case should not automatically include.

If the business was at a trough, and if margins could normalize, and if the takeover price reflected strategic value, then a recovery scenario mattered.

Claude surfaced that intuition quickly.

That is useful.

But the question is where that intuition belongs.

It does not belong inside the base case without being labeled.

It belongs in a scenario.

Sharpe AI's base case was strong.

Its scenario grammar was not strong enough yet.

The Bear / Base / Bull band of $11.91–$18.94 did not properly capture a true cyclical recovery case. It moved the model around the evidence-led base case, but it did not express the actual bull thesis: margin normalization from a trough.

That matters.

A real finance team would want both: an evidence-led standalone base case, and a labeled recovery scenario that says, "What if margins return toward historical levels?"

So the honest read is:

Sharpe AI protected the base case correctly. The next product improvement is richer scenario grammar: margin reversion cases, revenue recovery cases, management-case overrides, cycle-normalized EBIT margins, explicit market-implied scenarios, and acquisition / control-premium framing.

If this were a live investment banking assignment, the right conclusion would not be: Sharpe says $15, Claude says $27–$37, market says $55. Pick one.

The right conclusion would be:

Standalone evidence supports roughly $15. A recovery thesis could support $27–$37. The takeover price at $55 likely reflects control value, strategic value, or deal dynamics beyond the standalone base case.

That is an actual finance memo.

And Sharpe AI makes that memo stronger because it gives you the floor.

Without a defensible floor, the recovery thesis floats.

Without a labeled recovery thesis, the base case gets contaminated.

Without separating the takeover price, the model starts explaining a number it should be challenging.

This is where Sharpe AI stops being a spreadsheet assistant and starts becoming a decision-support layer.

The three cases now tell one story.

Sharpe AI is not trying to be the most confident model in the room.

It is trying to make the model's logic inspectable.

Sometimes that means finding a structural bridge issue.

Sometimes it means proving the evidence underneath a similar answer.

And sometimes it means telling the model:

That is exactly what financial model control should do.

The most dangerous model is not the one that is obviously broken.

It is the one that makes a reasonable story sound like evidence.

This case showed why Sharpe AI needs to exist.

In high-stakes finance, a model should separate what the evidence supports, what the analyst believes, and what the market is currently pricing.

Those are not the same thing.

When they get mixed together, the output becomes persuasive but fragile.

When they are separated, the model becomes useful.