How to write a whitepaper: the argument, the evidence under each claim, and the section most of them omit

A whitepaper is read by people who did not commission it and have no obligation to finish it. That single fact decides how it has to be built. The reader arrives with a position, usually a settled one, and leaves either holding a different position or holding the same one with a new reason to distrust the document. There is no third outcome in which the piece is merely “informative”: information the reader already has persuades nobody, and information the reader cannot check is discarded as marketing.

Most documents that call themselves whitepapers fail at exactly that point. They are long, they are well typeset, they contain a diagram of an architecture, and they make no argument that could be wrong. A claim that cannot be wrong cannot be checked, and a claim that cannot be checked cannot change what anyone does. The page count is real work; the structure underneath it is often decoration.

This piece is about the structure, because the structure is the whole of the method. It covers what a claim is in a document of this kind, what has to sit under each one, how to order sections so that the order carries the argument rather than announcing it, and the one section that turns a whitepaper from an assertion into an instrument — the section that says what would make it wrong. It ends with what a 10,000-word draft costs to produce at measured rates, and with the decision to take before the first word is written.

A whitepaper is an argument that has to survive a reader

Three things have to be true of the document before anything else about it matters.

It has to make a claim. Not a topic — a claim. “Data residency in the Gulf” is a topic. “A team of eleven can hold residency compliance for a corpus of this size without a dedicated legal function, provided the corpus is text and the retention rule is enforced at ingest” is a claim: it asserts something, it names the conditions, and a reader could disagree with it and be right. The first is a heading. The second is a whitepaper.

Every load-bearing sentence in the document has to be traceable to evidence the reader could examine. Not a link to a landing page, not “industry sources suggest”: a document, a passage, a date. Where the evidence does not exist, the document says so in the place where the claim would have been, rather than filling the space with a fluent sentence that reads like knowledge.

And the argument has to be falsifiable in a way the document itself states. A reader who finishes the piece should know which observation would overturn it, what conditions it depends on, and when it stops being current. Those three items are not a caveat appended to the end. They are the difference between a document that can be used to decide and a document that can only be quoted back.

These three requirements do more work than any style rule, and they are the reason a whitepaper cannot be assembled from a template with the client’s name inserted. A template supplies headings. It cannot supply the argument, because the argument depends on what the evidence in this particular corpus supports, and that is discovered by reading the corpus.

Structure is the argument, not a sequence of headings

An outline that could be reordered without loss is not an outline. It is a list of subjects.

The test is mechanical and it takes a minute. Take the section titles and ask, of each one, what it establishes that the document would be missing without it. If the answer is “background”, the section is doing nothing: background a reader needs can be given in a paragraph where it is needed, and background a reader does not need is the most common reason a whitepaper is abandoned on page four. If the answer is “it makes the case for the next section”, the section is a move in the argument and belongs where it is. If the answer is “it was in the template”, delete it and see whether anything breaks.

Table 1 sets out the moves a whitepaper’s argument usually has to make. It is deliberately not a template with fixed titles or a fixed count. The moves appear in the order the argument requires, some appear twice, some are two sentences long and some are four pages, and a document with a move missing is weaker in a specific and predictable way.

Table 1. The moves a whitepaper’s argument has to make, the work each does, and the test that it has actually been made. Source: this publication’s drafting method for long-form work; the document parts are those listed for the studio’s Writer role.

Two of those moves are worth arguing with, and both arguments improve the document.

The first is what is already accepted. Most technical documents spend their opening pages re-establishing facts their reader supplied the brief for. The cost is not the words; it is the reader’s attention, spent on ground they already hold. A better opening establishes only the common ground the argument actually stands on, cites it in a sentence, and gets to the claim. Where a reader will doubt the common ground, that doubt is the argument’s first move and deserves its own section — but then it is a move, not a preamble.

The second is what it means for the reader. This is the move most often left as an implication. The evidence is presented, the reasoning is sound, and the document stops, on the assumption that a competent reader will draw the conclusion. Some will. The ones who will not are usually the ones with authority to act, because drawing the conclusion requires holding the whole document in mind at once, and they are reading it in the gaps of a Tuesday. Write the implication out. If it cannot be written as a sentence naming who acts and what they would do, the document has not finished its argument.

The claim register

Before the outline, a register. Not notes, not a bibliography — one row per claim, filled in before drafting and checked after.

The register is what makes the difference between a document that has sources and a document that has evidence. A source list records what was read. A register records what each claim rests on, and it fails visibly: a row with an empty evidence field is a claim the document cannot make, discovered at the desk rather than by a reader.

Table 2. The six fields of a claim register, what each holds, and the failure it prevents. Source: adapted from the citation entry specified on the pillar for the research and editing stations; the falsification field is this publication’s addition.

The fifth field is the one that earns the register its place in a document of this length, and it is the one no template asks for. Writing it forces a distinction the drafting mind resists: between a claim that has been tested and a claim that has been believed for long enough to feel tested. “Unit cost falls by a third at the second site” is a belief until someone writes down what would make it false — the volume threshold below which the fall does not happen, and the observation that would show the threshold was not reached. Once that sentence exists, the claim in the body has a boundary around it, and the boundary is what a serious reader looks for.

Two rules keep the register honest. A row with no evidence is not deleted; it is carried into the draft as a marked gap, because the gap is information about the argument and hiding it produces a document that reads better and says less. And a row whose evidence supports the claim partly is not rounded up to supported: partial support is recorded as partial, and the claim is narrowed to what the passage actually carries.

The evidence under each claim

Four kinds of material cannot carry a claim in a document of this kind, and each fails for a reason that will not improve with better tools.

A link nobody opened. A page that was never retrieved is not evidence; the claim was not checked against anything, and a reader who follows the link may find a sentence that says something else entirely. Where a figure matters, the document that holds it has to be in hand, not behind an address.

A figure with no period. Numbers describe a stretch of time, and a document that omits which stretch cannot be tested later. “The rate fell” is unanswerable; “the rate fell between 2021 and 2023” can be checked and can be found wrong. The as-of date is not a citation convention, it is what makes the claim falsifiable.

A summary standing in for the document. An abstract, a review, a trade write-up of a study is a report about a report. It can be cited as second-hand and it cannot be the authority for a number. The specific failure is worth naming because it is everywhere: a summary says a study found “a marked improvement”, and by the third document in the chain the improvement has become a percentage that appeared in none of them.

A passage that circles the point. A document can spend thirty pages discussing a subject and never state the claim the citation needs. It is a source for having discussed the subject and for nothing else, and the only test that settles it is the quote: put the sentence beside the claim and see whether it carries the weight.

A document with these four rules enforced has a property that is easy to underrate and hard to fake: a reader who doubts one sentence can open one document and find the passage. That is the whole of the promise. It is not a promise that the sentence is true.

None of this is a claim about what a good whitepaper reads like. A document can satisfy every one of these rules and still be dull, badly paced and too long; the two failures are independent, and a well-sourced document with no argument is as common as an argued one with no sources. What the rules buy is narrower and more useful: every sentence can be traced, and the trace survives somebody checking it.

The section most whitepapers omit

Nearly every whitepaper ends its argument with what the evidence shows. Very few state what would make them wrong. That omission is not an oversight; it is a decision, usually made halfway through drafting, on the reasonable-sounding ground that a document is supposed to persuade and a paragraph of doubt does not help.

It is the wrong decision, and the reason is that the section does three jobs nothing else in the document can do.

It converts the claim into a test. A reader who has been given the assumptions, the falsifying observations and the scope conditions can hold the document against their own situation and reach a conclusion the document did not hand them. That is what a technical reader wants, and it is what makes the piece quotable in a meeting where the author is not present.

It dates the document. An argument has an expiry: a regulation changes, a price moves, a corpus grows past the size at which the claim was made. Naming the expiry — a date, an event, a threshold — tells the reader when to come back, and tells them that the author expected the argument to be tested rather than accepted. A document with no expiry is either making a claim that cannot be wrong or pretending that nothing will change.

And it moves the burden of doubt. A reader’s objections are strongest when they are unstated, because they stay in the reader’s head where the document cannot answer them. Written into the document as the falsification section, the same objection becomes a line the reader can check off. The persuasive effect usually runs the other way from the one feared: a document that names its weak points reads as a document written by someone who looked for them.

Four parts, in this order, about a page for a document of ten thousand words.

The assumptions. The things the argument takes as given, stated so that a reader can say which one does not hold for them. An illustrative whitepaper might argue that a second operating site cuts unit cost by a third. Its assumptions are that volume holds at eighty per cent of the combined capacity, that the second site’s fixed costs are carried at the same rate as the first, and that the delivery lead time does not lengthen. Each is a sentence a reader can disagree with, and each disagreement is a productive one.

The falsifying observations. For each load-bearing claim, the observation that would settle it against the document. For the illustrative claim: unit cost at the second site, measured across four quarters, at volume below eighty per cent of capacity. If that measurement shows no fall, the claim is not weakened; it is wrong, and the document said in advance what would show it.

The scope conditions. Where the argument does not apply. Site size, sector, regulatory regime, corpus size, the stage of the business. A claim drawn from one setting and applied in another is the ordinary failure of this kind of document, and the scope paragraph is where the author draws the line rather than leaving the reader to draw it wrongly.

The expiry. The date or event at which the argument has to be re-tested, and the reason. A rule change coming into force, a price band that moves with a supplier’s tier, a corpus that will double when a second department is added. The expiry is a service to the reader: it tells them the document is a dated instrument rather than a permanent truth, which is the only honest claim any document of this kind can make.

Two cautions, both learned the hard way by anyone who has written the section and then watched it argued with.

The falsification section is not a disclaimer. It is not the paragraph where the author hedges every claim to nothing. “Results may vary” is not an assumption; it is an admission that the argument was not made. Each item in the section has to be specific enough to be testable, which means it names a measure, a threshold or a date.

And it is not the place for the claims the author was unsure about. Those belong in the register as gaps, marked in the body where they occur. The falsification section tests the argument the document actually makes; it does not carry the ones the document declined to make.

What the drafting costs

The rate is measured rather than listed: $0.00700 a call for deepseek-v4-flash on Base under exact with a short prompt, read from the live 402 challenge on 31 October 2026 outside the service’s weekday peak windows of 01:00–04:00 and 06:00–10:00 UTC. A long prompt, and the upto ceiling at max_tokens: 1000, is $0.00955. The $0.00875 that circulates as a catalogue figure for this endpoint is Qwen’s rate, not DeepSeek’s; it appears in this piece only where it is named as such, and no figure below uses it as an input.

An illustrative ten-thousand-word whitepaper, eight sections, built from a source pack of twenty-four documents, prices as follows.

Table 3. An illustrative ten-thousand-word whitepaper at the measured rate on Base. Source: live 402 handshake, 31 October 2026; call counts are planning assumptions stated below.

The addition is short: 24 + 40 + 16 = 80 calls, and 80 × $0.00700 is $0.56000. That is $0.05600 for each thousand words, $0.07000 a section, and $0.00700 a document read — the reading is the cheapest line in the table and the one that decides whether the other two are worth anything.

The forty-call drafting figure is the one the studio’s Writer role publishes for a long piece, and it covers the drafting station alone. The twenty-four reading calls belong to the research station and the sixteen red-team calls to the pass an editor makes over a finished draft; a whitepaper that skips either is cheaper by exactly the amount it skips. The role page’s figure and this one are not in conflict: 40 calls is $0.28000, and the full document is 80 calls and $0.56000.

Three comparisons bound it. At the metered ceiling of $0.00955 the same eighty calls cost $0.76400, which is $0.20400, or 36 per cent, more — the gap between a short prompt under exact and a long one, or a call that authorises a thousand output tokens. At Qwen’s $0.00875 the same eighty calls would cost $0.70000, a quarter more, which is the premium for a second model and not a price for this one. And 0.50 USDC, the opening balance credited to a workspace, buys seventy-one calls at $0.00700: those seventy-one cost $0.49700, the last nine cost $0.06300, and the balance is $0.06000 short of the document. A piece of this length is not something to start on the first credit without knowing that.

What the figure covers is drafting. The corpus — the client’s documents prepared as text rather than photographed pages, the standards and reports that carry the evidence — is not in it, because preparing a corpus is work rather than a call. Nor are the hours of the person who decides which claims matter, reads the red-team section, and signs the document. On a piece of this length those hours exceed the entire call bill by a wide margin, and no measurement of them is offered here, because none has been taken.

Where the method stops

Four limits, stated because a piece about evidence that overstated its own would be a poor advertisement for the method.

The method guarantees traceability, not truth. A register built on a corpus of well-documented errors produces a document in which every claim resolves to a passage that is wrong. Nothing in this chain will tell you otherwise: it establishes where a sentence came from, and whether the sentence is correct is a property of the corpus and of the person who assembled it.

Structure cannot supply an argument that the evidence does not support. Where the corpus is thin, the honest document is short. The method will make a thin argument legible as thin, which is useful and is not the same as making it strong — and a client who wants the length of a whitepaper over evidence that supports a memo has a problem no drafting arrangement can solve.

No figure is offered here for how many people read a whitepaper, how far into one they get, or what one is worth commercially. Those numbers are widely quoted and none has been measured by this publication, and a plausible statistic invented inside a document about the value of evidence would be the worst possible place to start.

And the falsification section is only as good as the author’s willingness to be checked. A section written to satisfy a house style — four headings, three sentences each, no threshold anywhere — is worse than no section at all, because it borrows the credibility of the practice without doing it.

The decision

Write the claim register before the outline, and write the falsification section before the draft.

That order is the whole of the advice. The register tells you whether the document can be written at all: if four or five rows cannot be filled, the piece is not a whitepaper yet, it is a subject in search of an argument, and the honest move is to narrow the claim until the rows fill. The falsification section, written early, tells you what the argument is worth: a claim whose falsifying observation you cannot name is one you have not tested, and writing it last means discovering that after ten thousand words are committed to it.

Then apply one test to the finished document before anyone outside sees it. Cover the body and read only the claims as the register states them. If the document says what the register says, in the order the argument requires, the structure has done its job. If the register has six claims and the document makes eleven, the extra five arrived during drafting, and every one of them is a sentence nobody can check.

Where this goes

• The pillar: x402.review/2026/10/31/content-and-knowledge/ — the four stations, the citation chain and the measured month
• The role that drafts it: studio.x402.press/roles/writer — the document parts, the worked cost and the refusals
• The role that checks it: studio.x402.press/roles/curator_editor — the register read into text, the matrix and the flagged items
• The service it buys: x402.press/services/deepseek/ — the canonical price, every chain
• The second model: x402.press/services/qwen/ — the entry tier at $0.00875 and the ladder above it
• The reviews: DeepSeek and Qwen — the measured handshakes and what they do not promise
• The price ladder: inference endpoints compared — three endpoints, the metering ceilings and the cost of escalation

Method note

The price was read from the live 402 challenge on Base on 31 October 2026, outside DeepSeek’s weekday peak windows of 01:00–04:00 and 06:00–10:00 UTC. The catalogue figure of $0.00875 was checked against the handshake, found to be Qwen’s rate, and is named as Qwen’s wherever it appears; no arithmetic in this piece uses it as DeepSeek’s price. The call counts in Table 3 are planning assumptions and are labelled as such; the forty-call drafting figure is the one published on the Writer role page and the eighty-call total adds the research and red-team passes to it. The illustrative whitepaper, its second-site claim and its assumptions are constructed for this piece, and no client, engagement or published document is described. No readership, engagement or conversion figure is offered, because none has been measured here.

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: whitepaper, writing, structure, evidence