A research brief is not a topic and it is not a request. It is the contract between the person who needs to know something and the person or the agent who will find out: it names the question, the decisions the answer will inform, what is inside the scope and what is not, which sources are allowed, what will be handed back, and by when. Six fields, and every sentence the researcher writes afterwards is either licensed by one of them or invented on the researcher’s own authority. Invented authority is what produces the confident paragraph nobody can check.
The ordinary failure is not laziness. “Look into what model calls cost” is a topic, and a researcher handed a topic has to choose the question, choose the decisions, choose the source set, choose the standard of evidence and choose the form of the answer — and then answer its own choices while pretending they came from the client. That is work done twice, at the wrong end of the process, by the party with the least information about the answer, which is the party who does not have the documents. A brief exists so that those five choices are made once, by the person who knows what the answer is for, before anybody starts reading.
This piece sets out the six fields, what each one has to contain for the brief to be worth the paper, and what fails each. It shows one brief filled in on a subject constructed for the purpose, prices it at the Researcher assembly’s call counts and the measured rate on the endpoint’s own 402 challenge, and shows what the same question costs when it is started from a one-line request instead. It closes with what a brief cannot do, because a contract that is taken for a guarantee is worse than no contract at all.
The six fields
A brief has six fields and no seventh. Anything else a client wants to say — tone, house style, the names of people who must not be mentioned — belongs in a covering note, because it changes how the answer is written rather than what the answer must be. The six are load-bearing in a specific way: remove any one and the researcher supplies it from inference, and inference is where a piece of research quietly becomes an opinion with footnotes.
Table 1 gives each field, what it holds, the test it has to pass and the failure the test catches. The tests are written as questions a second person can answer from the brief alone, without asking the client what they meant. That is the same standard this publication applies to a claim list: if a reviewer has to go back to the author to know whether the field was met, the field was not written.
Table 1. The six fields of a research brief, what each holds, the test it has to pass and what fails it. Constructed for this piece from the brief structure the Researcher station builds to.
| Field | What it holds | The test it has to pass | What fails the test |
|---|---|---|---|
| Question | One interrogative sentence naming the subject, the population and the period | A researcher could answer it wrongly, and a reader could tell | Two questions joined by “and”; a topic; a question whose answer is already decided in its wording |
| Decisions | The decisions the answer will inform, and the date each is taken | Every part of the question can be traced to a decision that needs it | “For information”; “background”; a decision already taken, where the research is really a search for support |
| Scope | The material included, the populations included, the period covered, and the boundaries stated as exclusions | A document that a reasonable person might include is explicitly in or out | A subject area instead of a boundary; no exclusions, which means no boundary was decided |
| Sources allowed | Which documents may carry a claim: types, issuers, dates, and whether secondary accounts count | Any claim in the answer can be checked against a class of document named in the brief | “Credible sources”; a reading list of URLs; nothing, which licenses whatever the search returns |
| Deliverable | The artefact, its sections, its length, its tables, and the status of each claim in it | The client can say whether what arrived is what was asked for | “A report”; a format with no sections; a length with no audience |
| Date | The date the answer is due, and the as-of date every figure in it must carry | The researcher can plan the reading, and the client can test the answer later | One date where two are needed; a date after the meeting the answer was for |
The two fields most often dropped are the second and the sixth, and they fail differently. A brief without decisions produces research that is accurate and useless: the answer is true, the question was interesting, and nothing was decided by it. A brief without an as-of date produces research that is correct on the day it is read and untestable afterwards, because a figure without a period cannot be checked against the year it was meant to describe. Neither failure announces itself. Both are invisible in a well-written answer, which is the reason they persist.
The question, written so it can be answered wrongly
A question earns its place in the brief when a researcher could get it wrong. “What do model calls cost?” can be answered wrongly: the model matters, the chain matters, the token count matters, and a figure lifted from one endpoint’s price list is wrong for another endpoint’s challenge. “What is the state of the market for model calls?” cannot be answered wrongly, because it cannot be answered at all: there is no state, there is a set of prices that move, and the researcher will choose which set to describe and call the choice a scope.
Three things belong in the sentence. The subject, named precisely enough that two researchers would read the same thing: not “AI costs” but “the per-call price of a short chat completion”. The population, meaning which instances count: which providers, which chains, which model tier. And the period: the week, the quarter, the date the figure must be true for. A question with all three is answerable and refutable. A question missing one of them hands the missing piece to the researcher, who will pick the most convenient instance, the most available provider and the most recent date, and none of those choices will be visible in the answer.
One question per brief is a rule rather than a preference, because the alternative is not two answers. A brief that asks what calls cost and whether a second model is worth attaching produces one answer — usually the first, because it is the easier reading — and a paragraph of reasoning about the second that carries no evidence at all. If both matter, write two briefs. The second is cheap: at the counts below, a brief is eleven paid calls.
The question also fixes what the brief refuses to promise. It does not promise that the answer will be the one the client hoped for, and where the documents do not settle the point the answer says so. A brief written as a request for confirmation is not a brief; it is a brief with the answer in it, and the researcher’s only honest response is to return it.
The decisions, which decide the scope
The decisions field is the one that makes the rest of the brief arguable. “Which endpoint to attach to a month of drafting work” tells the researcher that price per call matters, that the price has to be comparable across providers, and that the comparison has to be at the token counts the work will actually use. “The board asked about AI costs” tells the researcher nothing except that somebody senior is curious, and a curious reader is served by a summary of the most available figures — which is the definition of a brief that produces a press round-up.
The test in Table 1 is deliberately strict: every part of the question must trace to a decision that needs it. Applied honestly, it usually shortens the question. A client who wants to know the price, the latency, the quality and the vendor’s roadmap for four providers is asking four questions, and three of them will be answered thinly. The decision — attach one endpoint, for this work, this month — needs the price, at the counts, on the chain the wallet uses, from the challenge in front of the call. It does not need the roadmap.
There is a second use for the field, and it is the reason a brief is a contract rather than a form. When the answer arrives and the client disagrees with it, the decisions field is what makes the disagreement specific. Either a decision was described and the research did not serve it, which is the researcher’s failure; or the decision was not described, in which case the research served a decision nobody wrote down, which is the client’s. Without the field, every disagreement about a piece of research becomes a disagreement about quality, and neither party can win that.
Scope, sources and the two rules that make the corpus
Scope is a boundary, not a subject. “Model pricing” is a subject and includes everything; “per-call prices for chat completions on one chain, read from the challenge, in the week ending on the date of the answer” is a boundary, and a document either falls inside it or does not. The brief states the inclusions and the exclusions in the same breath, because an inclusion with no exclusion has not been decided — it has been assumed, and the assumption will be the researcher’s.
Sources allowed is the field that decides whether the answer is checkable. It names a class of document rather than a list of links: what type, from which issuer, dated within what window, and whether a secondary account may carry a claim. The Researcher station will not fetch the open web, and it will not cite a document it has not read, so a brief whose source field is a set of URLs is a brief asking for something the assembly cannot do; the documents have to arrive as documents. Where the source set is what the client holds, the brief says so plainly and the answer’s evidence base is the size of that set. Where the client expects the researcher to find documents, the brief has to name where they will come from, because nothing in the chain searches.
A secondary account is the field’s sharpest edge and it needs a decision in the brief, not in the draft. An abstract, a review, a summary of a report: each can be cited as second-hand, and none can be the authority for a figure. Where the brief allows secondary accounts to carry claims, the answer will be built at one remove and will say so at the point of each claim; where it forbids them, the answer will be shorter and will carry gaps where the primary document was not supplied. Both are legitimate briefs. A brief that says nothing on the question gets whichever is more convenient, and the client will not know which one they bought.
The two rules that make the corpus usable are the same here as anywhere in this chain. The register is built from documents read into their own text, not from addresses; and every entry carries the as-of date of the passage, not the date it was read. A brief can require both in a sentence, and it should, because a brief that requires them is a brief whose answer can be tested by somebody who was not in the room.
One brief, filled in
Table 2 is a brief written out in full on a question constructed for this piece. The subject is deliberately ordinary: a small team choosing an endpoint to attach to a month of drafting work. What the table shows is not the subject but the shape — six fields, each answered in a sentence or two, each answer specific enough that a researcher could disagree with it.
Table 2. A research brief on endpoint choice for a month of drafting work, written field by field. Constructed for this piece; no client, provider, wallet or month is described.
| Field | The brief as written |
|---|---|
| Question | What does one short chat completion cost, at the token counts a month of drafting actually uses, on the four endpoints this team’s wallet can pay on the chain it holds — and what does the month cost at each? |
| Decisions | One decision, taken on the fifteenth: which endpoint the drafting work is attached to for the following month. A second decision depends on the first and is taken the same day: whether the per-message ceiling is set at the exact price or above it. Both are recorded in the workspace, not carried in a meeting. |
| Scope | In: per-call prices for short chat completions, read from each endpoint’s own 402 challenge, on the chain the wallet uses. Out: latency, output quality, context limits, vendor roadmap, annual contracts, and any price read from a catalogue rather than a challenge. Out also: models above the flash tier, because the work will not use them. |
| Sources allowed | The 402 challenge each endpoint returns, read on the day of the answer, and the endpoint’s own published price for the same call. Catalogues may be used to check a figure and are not permitted to supply one. No secondary account may carry a price. Where a challenge and a catalogue disagree, the challenge is the price and the disagreement is reported. |
| Deliverable | One table: endpoint, model, scheme, atomic amount, price per call, and the ceiling at the token count used — with the month’s total at each endpoint beneath it. A second table reconstructs any endpoint whose challenge disagrees with its catalogue. Every figure carries the timestamp it was read. Prose is limited to what the tables do not carry; where a price changed during the read, the change is shown rather than averaged. |
| Date | Answer due on the fourteenth, one working day before the decision. Every price in it carries the as-of date of its challenge; no figure may carry an older date without saying so in the row. |
Read the row values against the tests in Table 1 and the brief holds on each. The question names subject, population and period, and a researcher could answer it wrongly by averaging two schemes or by reading one catalogue. The decisions are dated and depend on the answer. The scope carries exclusions, and the exclusions are the ones a researcher would otherwise have assumed into the work. The source rule is a class of document with a precedent for conflict, and it forbids the one substitution that would make the price look better — a catalogue figure for a challenge. The deliverable is a structure, not a format. The date is two dates doing two jobs.
What is missing from Table 2 is as deliberate as what is in it. There is no statement of the answer’s tone, no instruction to be thorough, and no request for a recommendation. Each of those would be an invitation to prose in a document whose deliverable is a table, and a recommendation would be a decision the client has just taken back off their own desk.
What the brief costs, and what a thin one costs
The Researcher assembly prices a brief at nine DeepSeek calls — six reading the source set, three drafting the brief, the report and the table — plus two optional Qwen cross-checks on a contested claim. Table 3 puts the measured Base rates against those counts.
Table 3. One research brief at the Researcher assembly’s planning counts: six reading calls, three drafting calls, two optional cross-checks, at the rates read from each endpoint’s 402 challenge on 31 October 2026. Counts are the role’s assumptions, not observations; the document parts are outside the total.
| Call | Unit price | Calls | Cost |
|---|---|---|---|
| DeepSeek — reading the source set, six calls a brief | $0.00700 | 6 | $0.04200 |
| DeepSeek — drafting the brief, the report and the table, three calls | $0.00700 | 3 | $0.02100 |
| Qwen — cross-check on a contested claim, two calls (optional) | $0.00875 | 2 | $0.01750 |
| Total | 11 | $0.08050 |
Eleven calls and $0.08050 is the whole of a brief at these counts. Without the cross-checks it is nine calls and $0.06300, and the cross-check is worth its $0.01750 only where the brief’s question is genuinely contested — a price that two endpoints disagree about, a claim that two documents state differently. A brief that settles nothing contested buys nothing with it.
Table 4 is the comparison that makes the brief worth writing, and it is a comparison between two starts on the same question rather than between two prices. The first start is Table 2. The second is the same question asked as a one-line request — “find out what model calls cost” — where the researcher reads the first source set it can assemble, discovers from the client that the work is drafting at a specific token count on one chain, and reads a second set.
Table 4. Two starts on one question: the brief written as a contract, and the same question asked as a one-line request, priced at the measured DeepSeek and Qwen rates. Constructed for this piece.
| Start | Reading calls | Drafting calls | Cross-checks | Total calls | Cost |
|---|---|---|---|---|---|
| Brief written as a contract (Table 2) | 6 | 3 | 2 | 11 | $0.08050 |
| One-line request, one re-brief | 12 | 3 | 2 | 17 | $0.12250 |
| Difference | 6 | 0 | 0 | 6 | $0.04200 |
Six calls, $0.04200, is the price of not writing five sentences: the second reading pass, which is $0.04200 on its own, was spent discovering a scope the brief would have stated. In percentage terms the one-line start is 52.2 per cent dearer than the contract, and the six calls are not the expensive part. The expensive part is that the first read set was assembled without a source rule, which means the answer’s evidence base was chosen by availability, and the client has no way to tell afterwards which of the two read sets the surviving figures came from.
The rest of the price is small enough to state plainly and worth stating because it bounds the exercise. At the metered ceiling of $0.00955, the nine DeepSeek calls cost $0.08595 instead of $0.06300 — $0.02295, or 36.4 per cent, more — which is the difference between a short prompt on the exact scheme and a call that authorises a thousand output tokens. The figure that circulates as DeepSeek’s price, $0.00875, would put the same nine calls at $0.07875: $0.01575 more, or 25 per cent, and it is Qwen’s figure rather than DeepSeek’s. The 0.50 USDC credited to a workspace as working capital is seventy-one calls at $0.00700, which is six briefs at eleven calls each, with five calls — $0.03500 — left over: a month of briefs at two a week, or an eighth of one month of the Curator / Editor’s eight-section pack at 96 calls.
The counts will be wrong for any particular brief, and the direction of the error is predictable. A source set of thirty documents costs roughly three times what a set of ten does to read, which moves the reading line and leaves the drafting line where it is. A brief with no contested claim drops both cross-checks. A brief that has to be re-scoped after the first read costs another six calls, which is why the re-brief in Table 4 is the ordinary failure rather than the exotic one — and why the six fields are worth ten minutes at the start of the work rather than an argument at the end of it.
What a brief cannot do
The limitation is structural, and it is the one that a client who has just read the sections above is most likely to forget.
A brief cannot supply a corpus. It can name which documents may carry a claim; it cannot make them exist. Where the evidence base is four documents and the question needs twelve, the answer will have gaps, and the brief’s job is to have said so before the reading started rather than to have licensed a wider claim in the hope that something turns up. The Researcher station reports the gaps and names the document that would close each one; a brief that forbids gaps is a brief that asked for a fluent answer, and fluency is available at any price.
A brief cannot make the sources accurate. It controls what may carry a claim and at what distance from the primary document, not whether the primary document is right. A register built from well-documented errors produces an answer that is faithful and wrong, and the brief has nothing to say about that — accuracy is a property of the corpus, and the corpus is the client’s.
A brief cannot decide what the answer means. It states the decisions the answer will inform, which is a long way from making them. The researcher returns a table with gaps marked; whether the price difference between two endpoints is worth a migration, whether the contested claim matters, whether the answer should change the decision — those are judgements taken with the answer in hand, by somebody accountable for them. And a brief cannot be retrospective. Once the reading has been done without a source rule, applying the rule afterwards does not restore the evidence base; it only relabels what was already read.
One more thing a brief cannot do is worth naming because it is often asked for. It cannot promise that the answer will be checkable by the client’s own lawyer, auditor or regulator. Those readers have standards a brief does not carry: they will want the primary documents, the passage quoted, and a person with standing to sign. A brief can make the answer traceable, which is the part the chain controls; it cannot make it compliant, which is a different instrument.
The decision
Write the six fields before the first call, and treat a field that cannot be filled as the answer to a question the client has not yet decided. If the decisions cannot be named, there is no brief and the research will not be used. If the exclusions cannot be written, the scope has not been chosen. If the source rule is a list of URLs, the corpus does not exist yet and preparing it is the first job rather than a detail of the purchase. If there is one date, there are two and the second is missing.
Then let the brief do the work it was written for. The researcher reads the source set, drafts the answer, marks every claim with its passage and its as-of date, and returns the gaps with the document each one needs. The client reads a document whose shape was decided in advance, disagrees with a row rather than with the piece, and takes the decision the brief was written to inform — at $0.08050 for the calls, which is less than the meeting the answer was for.
The agent’s part is narrow and worth stating. It reads what it is given, keeps the register, drafts the answer against the brief, and flags what the sources do not carry. It cannot choose the question, cannot widen the source set to make an answer possible, and cannot turn a brief that asks for confirmation into an answer that tests it. Those stay with the client, and the brief is written so that they are decisions somebody can make on paper.
Where this goes
- The pillar: content and knowledge: from a source you hold to a sentence you can defend — the four stations, the claim list and the verdict
- The role that answers a brief: studio.x402.press/roles/researcher — the register, the claim list, the cross-check and the three refusals
- The role that receives its register: studio.x402.press/roles/curator_editor — study material, the citation matrix, the deck and the pack-level pass
- Where the drafting is priced: studio.x402.press/roles/writer — one voice, numbered tables, model calls only
- The services: DeepSeek and Qwen — the canonical per-call prices on seven chains
- The reviews: DeepSeek and Qwen — the measured handshakes and what they do not promise
- The price ladder: inference endpoints compared — the three endpoints, the ceilings and the cost of escalation
- What happens to the register next: turning source material into articles — the register, the claim map and an outline that follows the evidence
- The shape of the answer: report structure for decision-makers — what was asked, what was found, what it means
Method note
The prices were read from the live 402 challenge on Base on 31 October 2026: 7000 atomic units under exact, which is $0.00700 for a short prompt on deepseek-v4-flash, and 9550 under upto at max_tokens: 1000, which is $0.00955. Qwen flash’s $0.00875 was read from its own challenge in the same way, at the same settings. The call counts — six reading, three drafting, two cross-checks a brief — are the Researcher station’s published planning assumptions, and they are assumptions rather than observations of a real brief.
The brief in Table 2, the question it answers and the one-line request in Table 4 are constructed for this piece. No client, provider, wallet or month is described, and no document, catalogue or price list is quoted. The comparison in Table 4 turns on a single re-read of six calls; a real re-brief where the source rule changes the read set twice would cost more, and this piece makes no claim about how often a brief is returned, because nothing here has measured that.
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.
