Turning source material into articles: the register, the claim map, and an outline that follows the evidence

Somebody hands over a folder — reports, papers, a spreadsheet, three interview transcripts, a brand guide — and asks for an article. The article they have in mind exists already, in outline, in the brief: it makes a point they believe, in a voice they like, at a length they have budgeted. The folder is supposed to support it.

Sometimes it does. More often it supports a smaller version of the point, or a differently shaped one, or a point about a neighbouring subject. That gap between the article the client wanted and the article the material supports is the whole of this work, and it is not a drafting problem. It is discovered by reading, recorded in a map, and settled before the outline is written — because once the outline exists, every subsequent decision is made in service of it, and a writer who has committed to a shape will find a way to fill it.

The method below is the one that keeps the gap visible instead of closing it with prose. It has four instruments: a register that records what was read, a map that records what each claim rests on, a decision about the voice the evidence can carry, and an outline ordered by the evidence rather than by the house style. It ends with the thing that actually happens on real work — what to do when the sources will not support the piece the client asked for.

What arrives, and what it is worth

Four things arrive with a brief, and only one of them is evidence.

The corpus. The documents themselves, in whatever state they were kept. This is the evidence, and it is the only item on the list that can carry a claim. Its quality decides the article’s ceiling: a corpus of press releases supports a piece about what companies said, not about what happened.

The brief. What the client wants the article to say, for whom, at what length, by when. This is not evidence, and it is the single most common source of unsupported claims in published work, because a brief written by someone who knows their business reads like a statement of fact. “Our customers consolidate suppliers after the first year” is a belief until a document says it.

The house style. The voice, the formatting, the conventions, the words the organisation uses for its own things. This is real and it constrains the surface of the piece. It has no bearing on what the piece can claim, and treating it as though it did produces the familiar artefact: an article that sounds exactly like the organisation and says only what the organisation already believed.

The expected angle. Usually implicit, occasionally in a subject line. It is worth writing down, because the angle is a hypothesis about what the evidence will show, and the work of the next hour is testing it.

The distinction between the first item and the other three is the discipline of the whole job. Everything after this point is an apparatus for keeping them apart.

The register: reading before quoting

The first artefact is a register, and it records documents rather than addresses.

One row per document: what it is, who issued it, the date its content was current, its format, and what it was used for. The as-of date is separate from the date it was read, because they answer different questions: one tells a later reader whether a figure described the period they care about, the other tells them when the folder was open.

Four kinds of item cannot enter the register as evidence, and the reasons survive any improvement in tooling.

A source that exists only as a link is unread. Nothing in this chain fetches the open web; the page was not opened, so nothing in it was checked. A link is a pointer to evidence and not evidence, and where the only copy of a document is behind a URL, the document has to arrive as a document.

An undated document cannot be placed in time, and a claim that cannot be placed in time cannot be re-tested. This is not pedantry about citations: a figure without a period is a figure a reader must go and re-derive before relying on it.

A secondary account — an abstract, a review, a trade write-up of somebody else’s report — is recorded as second-hand and never becomes the authority for a figure. Where the primary document is not in the folder, the claim is carried at one remove and says so. The failure this prevents is specific and common: a summary reporting that a study found “a marked improvement” becomes the citation for “improved by 20 per cent”, and that number was in nobody’s document.

And a passage that does not contain the claim is not a source for it, however much the document appears to agree. The register quotes. If the quote cannot be produced, the document discussed the subject and stated nothing.

The claim-to-source map

The second artefact is the one that does the work, and it is drawn before the outline.

A claim list is not a list of topics. It is one row per sentence the article will assert, with the evidence beside it, and it is filled in before drafting so that the empty rows are discovered at the desk rather than by a reader. Four statuses are available and the discipline is in refusing a fifth: verified, meaning a document in the register states it; partial, meaning a passage supports part of the claim and not the rest; unsupported, meaning nothing in the register states it; and missing, meaning a document was expected and did not arrive.

Table 1 is a filled map from a constructed corpus: sixteen documents on whether a four-day week affects retention in a professional services firm. It is shown filled because an empty template teaches nothing — the judgement is in the rows that are not verified.

Table 1. A filled claim-to-source map for an illustrative article, from a corpus of sixteen documents. Source: constructed for this piece; the sources named are illustrative artefacts, not real publications.

Eight rows, four of which cannot be used as drafted — one unsupported, one that is partial at best, and two missing — and that is a good map rather than a bad one. Those four rows are the useful output: they name what the article cannot say, which is the information the client needs before the outline is agreed.

Two habits keep the map honest.

Partial is a status, not a failure. The third row is the case. The survey supports a claim about what its respondents reported and not about retention in general, so the row is recorded as partial and the article carries the narrow sentence — the survey’s respondents reported — rather than the general claim the row was standing in for. The narrow version is checkable and, in this instance, more interesting.

Nothing is rounded up. A claim that is nearly supported, or supported by something adjacent, is unsupported. The temptation to grade it verified because it is probably true is the beginning of the end of the map’s value, and it is a temptation that arrives with a deadline attached.

The voice the evidence can carry

The word register does double duty in this work. The source register is the inventory of evidence; the register of the piece is its formality. They are different things and the first decides the second.

A house style specifies the surface: the words the organisation uses for its own products, the formatting, the avoidance of certain constructions, the length of a typical paragraph. Applied after the evidence has decided the order, it is a service. Applied before, it becomes a template the evidence is stretched to fit, and the stretching is where the unsupported claim gets in.

Three decisions belong to the writer rather than to the house style, and they are worth stating explicitly at the top of the draft.

Who the reader is, in a sentence concrete enough to exclude somebody. “A practice manager who runs a firm of thirty to eighty people, has read one trial report and will not read a second” excludes the academic reader and settles most subsequent questions.

What the reader already knows, and therefore what the article does not explain. Space is the scarcest thing in a piece of eighteen hundred words, and the commonest waste of it is background the reader supplied the brief for.

What the evidence can support being said about them. This is the decision that distinguishes a piece built on a corpus from a piece built on a brief: the article’s claims are the map’s verified rows, and the map is not adjusted to make the article more useful.

The practical reason to write those three sentences down is that they settle arguments later without re-reading anything. When a house-style edit proposes a sharper claim, the question is not whether the sharper claim is better writing. It is whether a row in the map supports it, and the answer takes ten seconds.

The outline that follows the evidence

An outline drawn from a house template is drawn from the wrong source, and the failure is structural rather than stylistic. The template says: context, three findings, implications, conclusion. The corpus says: one strong finding, one finding that is true only in a narrow case, and a large question nothing in the folder addresses. A writer working from the template will produce the number of sections the template asks for, and will fill the weak ones with the best available material, which is not the same as evidence.

The outline the evidence orders is drawn from the map, by sorting the rows.

The strong verified rows become the sections. They come first, weighted by their strength, because they are what the article can say.

The partial rows become a section of their own, or a qualified sentence in a stronger section, and their qualification appears in the sentence rather than in a footnote.

The unsupported and missing rows become two things: questions the article names and does not answer, and a covering note to the client explaining what would answer them.

Anything left over is cut. A row that is interesting and bears on nothing is the beginning of a digression, and digressions in short pieces are where word counts go to die.

Two structural consequences follow, and both are unwelcome to clients until they see the draft.

The order will not match the house template. A piece whose strongest evidence is about what trials measured will open with what trials measured, even if the template opens with the problem. The template can be applied to the shape afterwards — the headline, the standfirst, the formatting, the closing call — but not to the sequence, because the sequence is the argument.

Some sections will be short, and one may be a paragraph. This looks like unevenness and is in fact proportion: a section carries as much evidence as the map has for it. Padding a thin section to match the others is the single most reliable way to produce a piece that reads well and cannot be defended.

When the sources do not support the piece the client wanted

It is as much the ordinary case as the exception, and it is worth having the moves in advance because the conversation happens under time pressure.

Four moves, in order of preference.

Narrow the claim to what the sources carry, and say so in the piece. The claim in the map’s fourth row — partner billable hours fell in one trial — supports a sentence about one trial. The article says that, and the client gets a piece whose central sentence cannot be attacked. This is the best outcome and it is available more often than it is taken, because the narrowed claim is usually more interesting than the general one it replaces.

Re-source, and name what is needed. The two missing rows are not failures; they are a shopping list. The firm’s own attrition is 18 per cent needs the client’s HR data; sector attrition averages 14 per cent needs a published labour-market release that states it; client coverage was maintained needs the service records from two trial sites. Naming the document, the holder and the date by which it is needed converts an objection into a task, and the task is usually somebody else’s.

Deliver the piece the evidence supports, with the gap named in the piece and in the covering note. The article is publishable, it is honest, and it says in one sentence where its evidence stops. The covering note names the claim the client wanted, the reason it is not in the article, and the document that would put it there.

Return the brief, with the reason and the cost. Where the claim is the piece — where an article about a 20 per cent retention improvement is not an article at all if the improvement is unsupported — the honest outcome is a return. A return is a legitimate deliverable and not a failure of the process: it is the correct output of a map that has no verified rows under the load-bearing claim.

One sentence does most of the work in all four cases, and it is worth having it written before the conversation rather than composed during it:

> The material supports a smaller claim than the brief asks for. The article below makes the smaller one, because the larger one would rest on a survey with no method note; if you can send the trial data for the second site, the larger claim becomes available and the section can be rewritten in an hour.

That sentence does three things: it states the gap, it states the reason, and it states what would close it. It is not an apology, it offers no opinion about the client’s judgement, and it hands back a decision rather than a problem.

What none of the four moves does is publish the claim anyway. No part of this chain will attach the number the client prefers to a document that does not carry it, and the reason is not moral fastidiousness: an invented figure outlives every correction made to it, and the correction reaches a fraction of the readers the original did.

What the article costs

The rate is measured rather than listed. DeepSeek deepseek-v4-flash on Base under exact with a short prompt is $0.00700 a call, 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; the long-prompt amount and the upto ceiling at max_tokens: 1000 are both $0.00955. Qwen qwen3.8-flash on Base is $0.00875, a quarter more, and its upto ceiling at a 1,000-token allowance is the same $0.00875 — metering at that allowance costs nothing, which is the only place in this comparison where the two endpoints are not simply a quarter apart. The $0.00875 that circulates as a catalogue figure for DeepSeek is Qwen’s rate; it is named as Qwen’s wherever it appears here.

An illustrative article of 1,800 words from the sixteen-document corpus above, with twenty-four claims in the map, prices as follows.

Table 2. An illustrative article from a sixteen-document corpus, at the measured rates on Base. Source: live 402 handshake, 31 October 2026; call counts are planning assumptions stated below.

The addition is in two parts, because two prices are involved. The forty-seven DeepSeek calls are 16 + 6 + 3 + 4 + 8 + 3 + 4 + 3, and 47 × $0.00700 is $0.32900. The two Qwen cross-checks at $0.00875 are $0.01750. Together: 49 paid calls and $0.34650, which is $0.02166 for each document read, $0.01444 for each claim mapped, and $0.19250 for each thousand words of the article.

Three comparisons bound it, and the first one is the interesting one.

At the metered ceiling of $0.00955 the forty-seven DeepSeek calls cost $0.44885, and the two Qwen calls still cost $0.01750 because Qwen’s ceiling equals its exact amount at this allowance. The ceiling total is therefore $0.46635, which is $0.11985 more than the exact total — 34.6 per cent, not the 36 per cent the DeepSeek calls alone would produce. Attaching a second model that does not move under metering damps the ceiling.

Pricing all forty-nine calls at Qwen’s rate would give $0.42875, which is $0.08225 more than the actual bill, or 23.7 per cent: the premium applies to two calls out of forty-nine, and the article is a DeepSeek purchase with a cross-check attached rather than a mixed purchase at an average price. Treating the two rates as interchangeable overstates this article by about a fifth.

And 0.50 USDC buys seventy-one calls at the DeepSeek rate. The article spends $0.34650 of it and leaves $0.15350, which is twenty-one DeepSeek calls with change — enough for a second article’s reading, not its drafting.

What the figure covers is the reading, the mapping and the writing. The corpus is not in it and the corpus is the expensive part: sixteen documents prepared as text, dated and filed is work no call is charged for. Nor are the client’s hours, the person who decides which claims matter, or the reading of the covering note. On a piece of this length those are the largest items, and no measurement of them is offered here.

The Qwen cross-check is worth its price on exactly one condition, which is worth stating because it is easy to buy without it. Two calls argue the other side of a contested claim — in this article, whether the survey row can support any claim at all — and the argument is useful because it produces the sentence the client will read. Where nothing in the map is contested, the two calls buy nothing, and the article is $0.01750 cheaper for not making them.

Where the method stops

Four limits, stated because the alternative is a piece that overstates its own evidence.

The map guarantees traceability, not truth. A corpus of well-documented errors produces a flawless map and a wrong article, and nothing in this chain will tell you so: it establishes where a sentence came from, and whether it is right is a property of the corpus and of whoever assembled it.

It cannot reach outside the folder. None of these roles fetches the open web, so a claim that needs a live figure, a current price or today’s regulation cannot be verified here at all. A citation chain records what a document said on a date, which is the wrong instrument for a value that has to be read at the moment of publication.

It cannot make the narrower article as attractive as the broader one. Following the evidence costs the client something — the sharper claim, the cleaner number, the headline that would have travelled. That cost is real, this piece offers no figure for it, and pretending the honest article is always the more effective one would be a claim without a source.

And no readership, engagement or conversion figure appears anywhere in this piece, because none has been measured here. A plausible statistic about how articles perform would be the least defensible sentence in a document about the difference between a claim and a belief.

The decision

Before agreeing the brief, ask for the four documents that would carry the four claims the piece depends on.

That is the whole of the decision, and it is taken before any drafting. If the four arrive, the article is writable at the client’s intended strength and the rest is execution. If they do not, the conversation is about which claim the evidence can carry — held now, while the client still has room to send something, rather than after the draft, when the only options are a compromise and an argument.

Then apply one test at the end, which catches nearly everything: read the map and the finished article side by side, sentence by sentence, and check that no sentence asserts more than its row. Every claim in the article appears in the map, every map row carries a status, and every unsupported row is either cut or named as a question. If those three hold, the piece can be defended to a reader who checks — which is the only defence that survives publication.

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 reads the sources: studio.x402.press/roles/researcher — the source register, the synthesis table and the optional cross-check
• The role that drafts: 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 services: DeepSeek at $0.00700 and Qwen at $0.00875 — the canonical prices, every chain
• 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

Both prices were read from the live 402 challenges on Base on 31 October 2026, outside DeepSeek’s weekday peak windows of 01:00–04:00 and 06:00–10:00 UTC. DeepSeek quoted $0.00700 under exact for a short prompt and $0.00955 for the upto ceiling at max_tokens: 1000; Qwen quoted $0.00875 under exact and the same $0.00875 under upto at that allowance. The catalogue figure of $0.00875 was checked against the handshake, found to be Qwen’s rate, and is named as Qwen’s throughout. The call counts in Table 2 are planning assumptions, not a measurement of any particular article. The corpus, the sixteen documents, the trial report and the survey in Table 1 are constructed for this piece: no real publication is named, no client is described, and the covering-note sentence is a model of a reply rather than a quotation from a correspondence. Nothing in this piece reports how many people read an article, how far they get, or what it converts, because none of that 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: writing, sources, articles, citations