A discounted cash flow fails in two ways. The forecast can be wrong, which is a judgement about a business, or the structure can be wrong, which is a mechanical fault that no amount of forecasting skill repairs. The second failure is the more common and the easier to find, because it leaves a trace in the arithmetic: a free cash flow that does not rebuild from its components, a discount rate that does not match the cash flow it discounts, a bridge that does not reconcile, a terminal value carrying a share of the answer nobody noticed.
This piece is about the second failure. It sets out what a DCF is made of, the fourteen structural checks that catch the faults before anyone argues about the forecast, an illustrative model rebuilt from its components so the checks can be seen working, the sensitivity grid and the reverse test that turn the model from a figure into a range, the honest limit of a structural review, and the measured cost of building one.
The four blocks
Every DCF, however long, is four blocks arranged in a fixed order. The forecast block produces the free cash flow for the explicit years. The discount block produces the rate. The terminal block turns the last forecast year into a perpetuity. The bridge block turns the enterprise value into a claim on equity. Each block has inputs, arithmetic and an output, and each one can be wrong while the other three are right.
What makes the structure worth checking separately is that the errors are silent. A spreadsheet will not object to a discount rate built on a debt weight that does not match the debt in the bridge, or to a working capital change with the wrong sign, or to a terminal value calculated on a cash flow that includes an interest charge. Every one of those produces a number, and the number looks exactly like the output of a correct model.
The checklist
The fourteen checks below are the ones that catch structural faults rather than differences of opinion. They are written to be applied in order, because the early ones change what the later ones should say.
Table 1. Fourteen structural checks for a discounted cash flow, the fault each one catches, and how that fault shows in the output. Source: constructed for this piece from the component build in Table 2; the checks are structural, and none of them tests whether a forecast is plausible.
| # | Check | The fault it catches | How the fault shows |
|---|---|---|---|
| 1 | The cash flow measure matches the discount rate | Free cash flow to the firm discounted at a cost of equity, or the reverse | The rate is too high or too low for the flow by the whole of the debt benefit, and the error grows with leverage |
| 2 | Free cash flow is rebuilt from its components, not typed in | A cash flow line entered directly, with no audit trail | The figure cannot be re-derived, and no component can be challenged on its own |
| 3 | Tax is charged on operating profit, not on profit after interest | The interest tax shield counted twice | Enterprise value rises, and the double count is invisible unless the tax line is rebuilt |
| 4 | Depreciation and amortisation are added back but capital expenditure is deducted | Replacement capex omitted, or net investment double-counted | A business with capex below depreciation for years on end is a business assumed to shrink |
| 5 | The change in working capital carries the right sign | A working capital build treated as a source of cash | Cash flow is overstated by twice the working capital movement |
| 6 | The forecast period is long enough for the business to reach a steady state | A terminal value applied to a company still mid-transition | The terminal share rises above 75 per cent without anyone deciding that it should |
| 7 | The terminal growth rate is at or below long-run nominal growth | A perpetuity that outgrows the economy for ever | Terminal growth above nominal GDP growth produces an unbounded claim and a very large terminal value |
| 8 | The terminal value is discounted from the end of the last forecast year | Terminal value discounted one year short, or to the wrong date | A gain or loss of one year’s discounting, material at any rate above 10 per cent |
| 9 | The discount rate is built from the same capital structure as the bridge | A target weight in the rate and a book weight in the bridge | Two different companies are being valued in the same model |
| 10 | The discount rate’s inputs are dated and sourced | A risk-free rate from an unstated date, or a premium from memory | The rate cannot be reproduced, so the valuation cannot be re-derived by anyone |
| 11 | Mid-year convention applied consistently or not at all | Discounting the forecast mid-year and the terminal value at year end | The enterprise value acquires a fraction of a year of growth that the model does not admit |
| 12 | Every output reconciles: explicit value plus terminal value equals enterprise value | A hard-coded enterprise value with components that do not add to it | The bridge starts from a figure the forecast did not produce |
| 13 | The sensitivity grid moves the rate and the growth rate independently | A one-axis table presented as sensitivity | The answer looks more robust than it is, because the two assumptions were moved together |
| 14 | A reverse test reads the model backwards from a given value | A model that can only be read in the direction it was built | The assumptions are never tested against a value the valuer did not choose |
Checks 1, 2, 3 and 5 are the four that most repay a second look, and all four are visible in a single pass over the component build. The remaining ten are conventions that a reader can verify from the pack in a few minutes each, which is the point of writing them down: a model with a stated structure can be reviewed by someone who did not build it, and a model without one can only be admired.
The model, rebuilt from its components
The construction below is not a forecast of any company. It is a three-year model with every line derived at the same growth rate, so that each component can be checked against the revenue line and the free cash flow can be rebuilt by hand.
Table 2. An illustrative three-year free cash flow to the firm, built from its components. Operating margin, tax rate and the investment ratios are held flat, and revenue grows at 10.00 per cent a year. Constructed currency units. Components are rounded to the cent and totals are the sums of the rounded components. Source: constructed for this piece; it describes no company.
| Line | Year 1 | Year 2 | Year 3 |
|---|---|---|---|
| Revenue | 1,000.00 | 1,100.00 | 1,210.00 |
| EBIT at 12.00 per cent of revenue | 120.00 | 132.00 | 145.20 |
| Tax at 25.00 per cent | (30.00) | (33.00) | (36.30) |
| NOPAT | 90.00 | 99.00 | 108.90 |
| Add depreciation and amortisation, 5.50 per cent of revenue | 55.00 | 60.50 | 66.55 |
| Less capital expenditure, 7.00 per cent of revenue | (70.00) | (77.00) | (84.70) |
| Less increase in working capital | (15.00) | (16.50) | (18.15) |
| Free cash flow to the firm | 60.00 | 66.00 | 72.60 |
The arithmetic in that table is deliberately dull, and dull is the test. Year 1 rebuilds as 90.00 plus 55.00, less 70.00, less 15.00, which is 60.00. Year 2 rebuilds as 99.00 plus 60.50, less 77.00, less 16.50, which is 66.00. Year 3 rebuilds as 108.90 plus 66.55, less 84.70, less 18.15, which is 72.60. Every line moves at 10.00 per cent, so the cash flow moves at 10.00 per cent, and any reader who finds a year where it does not has found the fault.
The interesting check is the one the table passes narrowly. Capital expenditure at 7.00 per cent of revenue and depreciation at 5.50 per cent means the business invests more than it consumes each year, which is consistent with a business growing at 10.00 per cent. Reverse those two ratios and the model would still produce a positive cash flow, still discount to an enterprise value, and would describe a company that is growing its revenue while liquidating its asset base.
Discounting those flows at 12.00 per cent, with a terminal growth rate of 2.00 per cent, produces the following. The three present values are 53.57, 52.61 and 51.68, which sum to 157.86; the terminal value at the end of year 3 is 72.60 multiplied by 1.02 and divided by 0.10, which is 740.52, and its present value is 527.09. The enterprise value is 157.86 plus 527.09, which is 684.95.
Two facts about that figure follow from the structure rather than the forecast. The terminal value is 77.0 per cent of the enterprise value: more than three-quarters of the answer arrives after the explicit period, on a perpetuity assumption, in a model whose forecast covers three years of a business nobody has described. And the explicit period contributes 23.0 per cent, which is the part of the valuation an analyst can actually defend line by line.
The grid, and the reverse test
A single enterprise value from a single rate and a single growth assumption is a point on a surface. The surface is what should be published.
Table 3. The sensitivity grid on the model in Table 2: enterprise value at three discount rates and three terminal growth rates. Constructed for this piece. Source: the model in Table 2.
| Discount rate | Terminal growth 1.50% | Terminal growth 2.00% | Terminal growth 2.50% |
|---|---|---|---|
| 11.00% | 727.87 | 762.33 | 800.84 |
| 12.00% | 657.39 | 684.95 | 715.41 |
| 13.00% | 599.19 | 621.66 | 646.28 |
The corners of that grid are 599.19 and 800.84, which is a range of 33.7 per cent around a central figure of 684.95 — on the same cash flows, with the forecast untouched. That is the honest width of a DCF answer when two assumptions are allowed to move by one point and half a point respectively. A pack that publishes 684.95 and not the grid is publishing the middle of a surface and calling it a value.
The reverse test asks the question the grid cannot. Instead of choosing a discount rate and a growth rate and computing a value, it takes a value — an offer, a previous round’s price, a figure the client has in mind — and solves for the growth rate that value implies.
On the model in Table 2, an enterprise value of 500.00 implies a terminal growth rate of minus 2.696 per cent, which is minus 2.70 per cent to two decimals. The arithmetic is short: the explicit period contributes 157.86 whatever the terminal assumption, so the present value of the terminal value must be 342.14; that is a terminal value of 480.68 at the end of year 3, which is 72.60 multiplied by 0.97304 and divided by 0.14696. The value the client had in mind, in other words, is a value for a business whose cash flow declines for ever. That is a statement a reader can test against their own view of the business, and it is not a statement any forward-built model produces.
Where the checklist stops
The limit is worth stating before the checklist is adopted. Every check above tests the arithmetic, and none of them tests whether the forecast is plausible. A model can pass all fourteen and value the wrong business beautifully: the structure will be sound, the bridge will reconcile, the grid will publish a defensible range, and the revenue path underneath will be a wish with a growth rate attached.
Three consequences follow.
The checklist cannot supply the discount rate. It can insist that the rate is built from stated inputs on a stated date with a stated capital structure, and it can catch the mismatch between the rate’s weights and the bridge’s. It cannot tell you whether 12.00 per cent is the right rate for this company, because that is a judgement about risk and about what an investor would require, and reasonable valuers differ on it by a point or more.
The terminal share is a warning, not a threshold. The 75 per cent line in check 6 is a convention used here, not a rule; a company growing fast in a growing market can legitimately carry a terminal share above it, and a mature utility can sit well below. What the check buys is that the valuer has to look at the number and decide, rather than let it appear.
The structure does not make the output traceable to a source. Checks 2 and 10 make the model re-derivable, which is the property a reviewer needs, but re-derivable is not the same as correct. A tidy model built on management accounts that have been adjusted twice is still a model built on adjusted management accounts, and the adjustments belong in the pack beside the figures they move.
What the model costs
Building the model is a small number of engine calls and a larger number of drafting calls. The engine operation dcf is published at $0.006125 on Base: a base-case call and a reverse-mode call are $0.01225 together. Those prices are published rather than measured, because the analyst services are pending review and return no 402 quote today.
The drafting is measured. On Base, deepseek-v4-flash quoted $0.00700 for a short prompt at max_tokens: 1000 on 31 October 2026, and $0.00955 for a prompt long enough to carry a filing. A model build is short-prompt work: each call carries one component, one check or one paragraph, not a document.
One illustrative build: nine calls to write the structural checks against the model as built, at $0.00700, is $0.06300; twelve calls to derive and sanity-check the twelve line items, which is $0.08400; six calls to rewrite the components the reviewer’s questions changed, which is $0.04200; three calls to draft the note’s method, range and limitation sections, which is $0.02100. Thirty calls, $0.21000, or $0.00700 a call — $0.02333 for each of the nine checks if the cost of the whole build is shared across them.
Three comparisons bound the figure. At the metered ceiling of $0.00955 the same thirty calls cost $0.28650, which is $0.07650, or 36.4 per cent, more than the measured total. At the $0.00875 that circulates for this service they cost $0.26250, $0.05250 more, or 25 per cent above. And the 0.50 USDC opening balance is 71 calls at $0.00700, so a build of thirty calls leaves 41 calls in the balance, which is a second model and most of a third.
What the total does not include is the reading. A model built from a filing rather than a summary puts the filing in the prompt, which moves those calls to the long-prompt rate or to Qwen, and the paragraph above is priced for a build from material already normalised. The document parts — extract, report, data table, office — carry no published per-call price and are outside every figure here.
The decision
Two practices turn a model into something a second person can review, and both are cheap.
Publish the component build beside the output. The table in this piece is a page of arithmetic that lets a reader rebuild the cash flow from revenue in three lines; a model that publishes only the enterprise value and the assumptions leaves the reviewer with nothing to check. If the components cannot be published because they were never built, that is the finding, and it is worth having before the number is used.
Then run the reverse test before the forward one is presented. Take the value the reader already has in mind and solve for the growth it implies. It is one call and one line of arithmetic, and it converts the exercise from a valuation into a comparison between two sets of assumptions — the valuer’s and the reader’s — which is the only form in which the disagreement can be discussed.
Where this goes
- The pillar: x402.review/2026/10/31/finance-and-markets/ — three engines, four jobs, one price
- The role: studio.x402.press/roles/financial_analyst — the agent that runs the DCF operation and writes the note
- The service it buys: x402.press/services/deepseek/ — the canonical price, per model and per chain
- The review: x402.review/2026/10/31/deepseek-review/ — the measured handshake, the prompt bands and the ceiling
- The cost context: x402.review/2026/10/31/inference-endpoints-compared/ — what length does to the bill, measured
Method note
The model prices were read from the live 402 challenge on Base on 31 October 2026 at 08:50 UTC, outside DeepSeek’s weekday peak windows of 01:00–04:00 and 06:00–10:00 UTC. The dcf engine price of $0.006125 is the published catalogue figure; the analyst services are pending review and return no 402 quote, so it is published rather than measured. Call counts are planning assumptions stated in the cost section. The revenue path, margins, investment ratios, discount rate, terminal growth rate and every value in Tables 2 and 3 are constructed for this piece and describe no company; the 75 per cent terminal-share line is a convention used here, not a rule. Totals are the sums of the rounded components, so a recomputation from unrounded values may differ by a cent, and the reverse-test growth rate of minus 2.70 per cent is a solution to the model in Table 2 rounded to two decimal places. Nothing here is investment, legal or tax advice, a valuation of any company, or a recommendation to buy, sell or hold anything.
x402.review is written and edited by N.P. Vincent. The publication is connected to x402.press, which sells the services reviewed here; that interest is disclosed on every piece that touches a service, and it is stated here rather than left to inference.
Filed under Jobs · Tags: dcf, modelling, checklist, valuation
