A charge at the end of the window

On the close of Friday 18 September 2026, Cognizant traded at $59.87. Sperio values a business from its filed accounts and a discount rate, giving no credit of any kind to future growth, and arrives at $101.39.

The market is paying 59 cents for each dollar the accounts support without growth. That is not among the twenty widest gaps in the index this week, and it is not the name our machine put first. It is the second cheapest of the 104 companies that clear our two gates — and our own screen blocks it. Explaining why we publish it anyway is most of this article, and the explanation is an admission: this is an editorial choice made against a filter we built, on a reading of the accounts that the filter cannot make.

Why not the widest discount, and why not the machine's pick

The widest discount in the index this week belongs, again, to a fertiliser company whose net income is down 85%. Behind it sit a building-products distributor at −84% and a laboratory landlord. Ranking by discount answers "what looks cheap", and it answers it reliably with whichever business is deteriorating fastest, because a collapsing denominator makes any price look like a bargain.

So the ranking applies two hard gates first. Does the business durably earn more on its capital than the capital costs — return on invested capital clearing ten percent in four fifths of the readable fiscal years and in each of the last three. And are both lines going the right way — revenue and operating income each compounding upward over the full readable history. Of the 455 companies the board has valued, 104 clear both.

The survivors are then ordered by cheapness: enterprise value to the median operating profit of up to fifteen fiscal years. That ordering is the only one in this project with any evidence behind it — a small backtest, an upper bound, described in last week's piece and not repeated here. The list this morning, cheapest first:

A. O. Smith was last Sunday's subject. The machine put it first again this morning, and the honest reading of that is that nothing has changed in a week — which is what the scoring section at the end of this piece reports. The next name on the list is Cognizant, and it is blocked by two flags on one line: over the ranking's four-year window, fiscal 2022 to fiscal 2025, net income is down 3% while earnings per share are up 3%, because the share count fell 6%. The screen reads a per-share line lifted by buybacks on a business that earned less. That flag was written three weeks ago, after a health insurer printed +83% on earnings per share over a net income down 11%, and there it was exactly right. Here we think it is wrong, and the rest of this is why.

What it earns

Return on invested capital, across the readable record: about 19%. Return on equity 14.9%, gross margin 33.4%, and more cash than debt — $1.9bn of cash against a $0.6bn term loan at the last balance sheet date. And the line that matters most for a valuation built on cash: in fiscal 2025 the company reported $2.23bn of net income and $2.60bn of free cash flow. It collected more than it earned.

The accounts, five fiscal years, as filed in the 10-K:

Revenue is up 14% over the five years, with one flat year in the middle. Operating income is up 20%, and the two low years in the middle each carry a restructuring charge — $229m in 2023, $134m in 2024, from a programme the company called NextGen. Fiscal 2025 carried no charge and a $62m gain on the sale of property; as filed, income from operations rose 17.2% on the year. On the company's own adjusted figure, which strips out both the gain and the prior year's charge, it rose 9.9%. Take either. They point the same way, and neither is a flat line.

The charge at the end of the window

Measured from fiscal 2022, the start of the ranking's window, earnings per share are up 3%, net income down 3%, the share count down 6%. Read like that, all of the per-share growth is arithmetic, and that is exactly what the screen read.

But one year does not fit, and this time it is the last one. In fiscal 2025, income before tax rose 18%, from $2.94bn to $3.48bn. Net income was flat. The whole gap is on one line: the provision for income taxes rose from $713m to $1,258m, and $390m of that is a single item. In the third quarter of 2025 the company wrote off a deferred tax asset of that size after a change in United States tax law — the repeal of the requirement to capitalise research costs meant the asset could no longer be used. The 10-K calls it a one-time, non-cash income tax expense, puts its effect on diluted earnings per share at $0.80, and says that beyond it the new law is not expected to move the tax rate.

Take that charge out and the window reads differently: net income up 14% from fiscal 2022 instead of down 3%, earnings per share up 21% instead of up 3%. The 6% fall in the share count is still there, and it still flatters the per-share line — but on a business that earned more, not less.

Last week's piece took a number away from our own output because a one-off charge at the start of the window made A. O. Smith's growth look three times larger than it was. This week a one-off charge at the end of the window makes Cognizant's growth look like zero. Same screen, opposite error. The screen is a filter, not a thesis: it removes traps, and on a four-year window with one non-recurring item in it, it will sometimes remove a company instead. That is the case for having an editor on top of the machine, and it is also the case for saying, in print, when the editor overrules it.

The downside, priced

A single valuation cannot be argued with. Three can.

The trough case sits below the price. At $46.97 against the $59.87 the market is paying, a buyer who is wrong about everything except the trough is down about a fifth, not protected. That is disclosed on the row, next to the two flags, and it is the reason the credited figure on the third line should be read as the frame's arithmetic and not as a target.

The rate sweep says the same thing from the other side. At today's cost of capital — 9.94%, built on the 4.94% ten-year Treasury as FRED reported it on 18 September — 104 companies clear both gates and five of them are worth more than their price even in the trough case. Raise the cost of capital by a hundred basis points and 89 still clear the gates but none keeps that cushion; at two hundred, 77 and none. Cognizant is not among the five at any rate.

What the price is saying

Something more violent than a flat line bending. In January the market paid about $85 for this company. At the end of June, $38.50. On Friday, $59.87 — thirty percent below the high, fifty-five percent above the low. In between, the accounts for the first half of 2026 printed revenue up 5% and income from operations up. The market changed its mind twice in eight months about a business whose numbers barely moved.

There is some support for the market's case in the most recent quarter. Revenue grew 4.5%, not the 7% of the fiscal year. In that same quarter the company opened a new restructuring programme — Project Leap, "aimed at streamlining operations and enhancing productivity through AI-led efficiencies", in the words of its 10-Q — and booked $84m against it. A business that restructures itself in the name of artificial intelligence is making the market's argument and its own at the same time: the market says machines will do this work for less, and the company says it intends to be the one that owns the machines.

The case against, in the board's own words

Sperio runs a devil's advocate whose only job is to attack the thesis. It is reproduced without softening, because a recommendation that publishes only its bull case is one that cannot be retracted honestly later:

Severe competitive erosion risk: the transition to AI-driven delivery models threatens to commoditize Cognizant's traditional labor-arbitrage moat if 'Project Leap' fails to achieve structural efficiency.

And from the same report's bear case:

A recent 37% rally leaves the stock vulnerable to a correction if upcoming earnings do not validate the AI-pivot narrative.

That is a real argument, and the accounts cannot answer it. They describe what the business earned; they say nothing about what a machine will charge for the same work next year, or whether the clients who pay $21bn a year for it will keep paying a company in the middle. What the accounts can say is narrower. A business at a 19% return on capital, with rising revenue and more cash than debt, is worth about 1.7 times its price on a no-growth valuation and about four fifths of its price at the trough. The distance between those two numbers is the distance between "this is a business" and "this was a business", and nothing in the filed figures can settle it.

Why now

Two things, and neither is a forecast. The market has repriced this company twice in eight months on accounts that moved a few percent — that is a fact about the price, not a prediction about it. And the number that blocks the name in our own screen is a tax charge that stops being the last line of the window on the day the next annual accounts are filed. We would rather say that than manufacture a catalyst.

One date is worth watching. The next results are scheduled for 28 October 2026, with consensus at $1.46 a share. For context and not as a prediction: the last four quarters came in against consensus at +7%, +2%, +5% and −1%. The number to read on that day is not the beat or the miss. It is revenue growth — whether it is still four to five percent, or whether what the market priced in June has started to show up in the accounts.

Last week's pick, scored

Last Sunday's piece said that this one would report whether A. O. Smith still clears the gates, and score it out loud.

It clears them, and this morning the machine put it first again: sixth cheapest of 104 survivors and the first without a blocking flag, at 15.3 times full-cycle EV/EBIT, valued at $121.13 with no growth credited. Its shares closed Friday at $56.95 against the $57.39 of the previous Friday, down 0.8% in a week — a number that says nothing about the thesis and is reported because it was promised. So why is this piece not about A. O. Smith again? Because it was published last week, and because the next name on the list is blocked by a reading of the accounts we think is wrong. That is an editorial choice, and a reader is owed the sentence that says so.

What this is, and is not

Every figure here comes from accounts filed with the SEC — the 10-K for fiscal 2025, filed on 12 February 2026, and the 10-Q for the quarter to 30 June 2026 — from a board valuation dated 20 September 2026, and from a ranking run dated 20 September 2026. The price used is the close of Friday 18 September 2026. The cost of capital is built from FRED observations as of 18 September 2026. One note on the company page on this site: for fiscal 2023 to 2025 it currently shows the company's adjusted income from operations rather than the figure as filed, a data-source mix we are correcting; the table above uses the 10-K.

The valuation gives no credit to growth beyond what the franchise record earns, and the ranking that produced this list orders the survivors of two gates by their full-cycle operating multiple — a way of comparing them to each other, tested on a small sample, not a forecast of what any of them will do. This week the editor overruled the machine on one name, for a reason stated above; the machine's own pick is reported next to it so that a reader can disagree.

The board that produced these figures will run on any company you name. What it will not do is decide for you — nothing here is investment advice.