A literature review is a claim about a body of work: that these documents exist, that they say these things, and that between them they settle — or fail to settle — a question. The claim is defensible only to the extent that the way the body was assembled can be inspected. Not the way it was summarised, which is prose, but the way it was found, screened and admitted: the searches that produced the records, the rule that decided which records were read in full, the rule that decided which were counted, and the statement of what the search could not see.
Most reviews fail on the first of those and are defended on the third. A reviewer reads whatever is to hand, writes a synthesis, and compiles a reference list that looks like a body of work; the reference list is then offered as evidence that the body was assembled deliberately. It is not. A list of documents is the output of a search, and a search that was not recorded cannot be distinguished from a search that was unrepresentative, except by reading the list — which nobody does, because the list is long and the prose is fluent.
Without database access, none of this changes; only the claims shrink. The four artefacts below are what make a review built from publicly reachable material defensible: a search record, an inclusion rule, an exclusion rule and a gap statement. Each is small enough to write in a page, each is checkable by a second person, and each exists to make a specific overclaim impossible. What they cannot do is make the review comprehensive, and the last section says why at length, because that is the claim a review without a library is most likely to be sold on.
One constraint has to be stated before the artefacts, because it decides who does what. Nothing in the research or editing assemblies fetches the open web: they read, write, render and mail. The searches in this piece are therefore run by a person in whatever public interfaces they can reach, and the documents that come back are handed to the workspace as files. That is not a limitation of the method — it is the method’s first recorded fact, and it belongs in the search record beside the query strings.
The four artefacts
Table 1 gives each artefact, what it holds, the test it has to pass and what fails it. The tests are written so that a second person can apply them to the review without the reviewer present, which is the only sense in which a review is checkable at all.
Table 1. The four artefacts of a defensible review, what each holds, and the test it has to pass. Constructed for this piece from the register and claim-list structure the research station builds to.
| Artefact | What it holds | The test it has to pass | What fails the test |
|---|---|---|---|
| Search record | Every search: the interface, the query string as typed, the date, the filters, and how many records came back | Another person could re-run the search and get substantially the same records | “A search of the literature”; a list of databases with no queries; searches logged after the results were read, so that the queries were written from the answers |
| Inclusion rule | The conditions a record must meet to be counted, written before screening | Every included record can be shown to meet every condition, from the record itself | A rule written after the reading, describing what was read; “relevant to the question”, which is a feeling with a citation |
| Exclusion rule | The reasons a record was dropped, each with the stage at which it was dropped, and the count against each reason | Every dropped record has one reason and one stage, and the counts reconcile with the funnel | A single exclusion bucket; reasons that are really the inclusion rule negated, with no counts; records that vanish between the search and the reference list |
| Gap statement | What the search could not reach, what literatures are known to exist behind it, and what evidence would change the conclusion | A reader can name the review’s blind spots after reading it once | Silence, which reads as completeness; “further research is needed”, which names no literature and no gap |
The four are ordered by when they are written, and the order is the discipline. The search record is written while searching, not afterwards, because a query reconstructed from the results is a query written to retrieve what the reviewer already read. The inclusion rule is written before the first record is screened, because a rule written afterwards describes the sample rather than selecting it. The exclusion rule is written as a log, one line per dropped record, because the funnel’s arithmetic is what shows the review was screened rather than skimmed. And the gap statement is written last, from the record rather than from memory.
The search record
A search record has five fields and no interpretation. The interface, because a general web search and a preprint server return different worlds. The query string exactly as typed, because “invoice matching” and “invoice matching automation” are different searches and the difference decides which literature appears. The date, because a review is a snapshot and the interfaces move. The filters — date limits, language, document type — because a filter is an exclusion applied before anyone sees the record, and an unrecorded filter is an exclusion with no line in the log. And the number of records returned, which is the denominator for everything downstream.
Two habits make the record usable. The first is to write the line before looking at the results, so that a query that returns nothing is still recorded: a search that found nothing is evidence about the search, and it is the only evidence the gap statement will have. The second is to record the queries in the form they were typed rather than tidied, because a reader re-running the search needs the string, not a description of it.
What the record buys is not rigour in the abstract. It buys the ability to say what the search was, which is the difference between “a review of the literature” and “eight searches across four open interfaces on 31 October 2026, listed in the appendix, returning ninety-six records”. The second sentence is weaker-sounding and far stronger: it can be tested, disputed and extended, and a reader who knows the field can look at the queries and say immediately which literature is missing, which is exactly the conversation a review is supposed to start.
The inclusion rule, and the exclusion rule beside it
An inclusion rule is a set of conditions that a record must meet to be counted, and the conditions have to be checkable from the record rather than from the reviewer’s judgement. Five conditions do most of the work: the record addresses the question’s population; it reports something — a finding, a count, a method, an argument — rather than announcing something; it is dated inside the review’s period, or its date is stated where it is outside; it is retrievable in full, not only as an abstract; and it is either a primary account or an explicitly secondary one, named as such.
The last two conditions are where a review without a library is decided, because they are the ones a paywall turns into a silent exclusion. A document that is retrievable only as an abstract cannot carry a claim about what it found; a document that is secondary can carry the claim at one remove and must be marked so. A review that includes both without distinguishing them is a review whose reference list overstates its evidence base, and the overstatement is invisible because every entry is a real document.
The exclusion rule is the inclusion rule’s log rather than its mirror. Each dropped record gets one reason and the stage at which it was dropped, and the counts have to reconcile with the funnel: records found, minus duplicates, minus those excluded on title and abstract, minus those excluded at full text, equals those included. Table 2 is that funnel for a constructed review, and the arithmetic is the point of it.
Table 2. The screening funnel for a constructed review: eight searches, ninety-six records, twenty-five included. The counts are constructed for this piece; the reconciliation between stages is the requirement.
| Stage | Records | What leaves the stage | Remaining |
|---|---|---|---|
| Records returned by eight searches | 96 | 23 duplicates across interfaces | 73 |
| Screened on title and abstract | 73 | 31 excluded: off the question, or no report of any kind | 42 |
| Read in full | 42 | 17 excluded against the inclusion rule | 25 |
| Included in the review | 25 | — | 25 |
Twenty-five of ninety-six records is 26.0 per cent, and the two intermediate shares are worth stating because they are what a reader uses to judge whether the screening was real. Fifty-seven and a half per cent of the screened set reached full text, which is high for a broad question and low for a narrow one, and 40.5 per cent of the full-text set failed the inclusion rule — a share that says the title-and-abstract screen was doing its job rather than passing everything through.
Table 3 is the exclusion log for the second stage, and it is deliberately tedious. Four reasons, seventeen records, one line each in the workspace register, and no bucket called “other”.
Table 3. The exclusion log at full text: four reasons, seventeen records, counts reconciling with Table 2. The reasons are constructed for this piece and are the ones a question of this shape produces.
| Reason | What it means | Records |
|---|---|---|
| Off the question | Addresses a different population, period or operation than the one the question names | 6 |
| No method stated | Reports a conclusion with no account of how it was reached, or of what was counted | 5 |
| Superseded | An earlier version of a dataset, standard or count that a later included document replaces | 4 |
| Not retrievable in full | Only an abstract or a landing page is reachable; the text cannot be read | 2 |
| Total | 17 |
The fourth reason is the one a review with database access rarely has to write down and one without it must. Two records left the funnel because their full text was not reachable, and those two are not failures of the review — they are the most useful lines in it. They belong in the gap statement by name, with the claim each would have carried, so that a reader who does have access knows exactly which two documents would change the picture.
The gap statement
The gap statement is the part of a review that a reader should read first, and it is the shortest. It says four things, in this order: which interfaces were searched and which were not; which literatures the reviewer knows to exist behind those interfaces and could not reach; what kind of evidence would change the review’s conclusion; and what the review therefore does and does not establish.
The first item is a list, and its value is in the negative half. “Searched: four open interfaces, one preprint server, two regulator sites. Not searched: any subscription database, any publisher’s full-text collection, and any literature in a language other than English.” That sentence tells a reader where the blind spots are, and it is a sentence most reviews never write because it concedes something. It concedes nothing that the reference list does not already imply; it merely stops the implication from being drawn by the reader in the wrong direction.
The second item needs knowledge rather than effort: a reviewer who knows the field can name the journals, the standards bodies and the working-paper series that publish on this subject and are not open. Where that knowledge is absent, the honest entry is that the review cannot say what it could not reach — which is itself a finding about the review’s scope, and better stated than implied.
The third item is the one that makes the review useful to somebody else. “A single primary study of the practice in this sector, with a stated sample, would move the claim from one remove to verified; two would settle it.” A reader can act on that sentence, and a reviewer who cannot write it has not understood what their own review rests on.
The fourth item is a sentence rather than a paragraph, and it is the only place in the review where the reviewer states their own ceiling. “This review establishes what the open literature reports and how consistently it reports it. It does not establish what is true in the sector, and it cannot say that no other evidence exists.”
One review, costed
Table 4 prices the four artefacts at the research station’s own call counts and the measured rates on Base: eight calls to screen seventy-three records in batches of ten, a call for each of the forty-two full texts, six calls to draft the synthesis from the twenty-five included sources, and two Qwen cross-checks where the synthesis turns on a contested reading.
Table 4. A constructed review at measured Base rates: screening, full-text reading, synthesis and a contested-reading cross-check. Counts are planning assumptions; the document parts carry no published per-call price and sit outside the total.
| Step | Calls | Unit price | Cost |
|---|---|---|---|
| Screening seventy-three records, ten to a call | 8 | $0.00700 | $0.05600 |
| Reading forty-two full texts, one call each | 42 | $0.00700 | $0.29400 |
| Drafting the synthesis from twenty-five included sources | 6 | $0.00700 | $0.04200 |
| Qwen cross-check on the contested reading | 2 | $0.00875 | $0.01750 |
| Total | 58 | $0.40950 |
Fifty-eight calls and $0.40950 is the whole cost of the review’s evidence handling at these counts, and three things about it are worth stating because they are what a buyer misjudges. Screening is cheap: eight calls, $0.05600, or $0.00077 a record. Reading is not: forty-two calls and $0.29400 is 71.8 per cent of the total, because every document that reaches full text is read rather than skimmed. And the synthesis is small — six calls, $0.04200 — because by then the work has been done and the drafting is the shortest step in the chain.
The third of those is the one that changes how a review is bought. On a corpus of twenty-five included sources the reading is $0.29400 and the drafting is $0.04200, a ratio of seven to one. A reviewer who tries to save money by narrowing the reading and widening the synthesis produces the ordinary failure: a confident overview over a thin corpus, priced at almost nothing and worth the same.
The bounds are the ones this publication uses everywhere, and they are small enough to state in two sentences. At the metered ceiling of $0.00955 the fifty-six DeepSeek calls cost $0.53480 instead of $0.39200 — $0.14280, or 36.4 per cent, more — which is the difference between a short prompt and a call authorising a thousand output tokens. The figure that circulates as DeepSeek’s price, $0.00875, would put the same fifty-six calls at $0.49000, $0.09800 or 25 per cent more, and it is Qwen’s figure rather than DeepSeek’s. The 0.50 USDC credited to a workspace is seventy-one calls at $0.00700: this review fits inside it with thirteen calls, $0.09100, to spare.
What a review without a library cannot claim
The limitation is not one limitation but five, and they are worth naming separately because each is overclaimed in a different kind of review.
It cannot claim completeness, or anything that sounds like it. “The literature”, “the available evidence”, “the body of work on this subject”: every one of those phrases is a claim that the search reached the edge of a field, and no review built from open interfaces knows where the edge is. What it can claim is what it searched and what it found, which is a weaker and more useful sentence.
It cannot claim that nothing exists. The absence of a document in an open corpus is not evidence of absence in the field; it is evidence that the document is not open, not indexed in the interfaces used, or not in the language searched. A review that writes “no studies were found” over a search of four open interfaces has made a claim about the world out of a fact about its own reach.
It cannot claim systematic-review status. Systematic reviews carry apparatus this method does not have: a protocol registered before the search, dual independent screening, a risk-of-bias instrument applied to every included study, and a quantitative synthesis where the studies permit one. A review can borrow the discipline of a recorded search without borrowing the name, and borrowing the name without the apparatus is the one thing in this piece that could fairly be called a misrepresentation.
It cannot quantify publication bias, and it is more exposed to it than a review with database access. The literature that is easy to reach is the literature that was published, and published literature over-represents findings over null results. A review of open material therefore inherits that skew without the means to measure it — which is a sentence for the gap statement, not a reason to stop.
And it cannot check the sources it does reach. Screening decides whether a document meets a rule; it says nothing about whether the document is right. A review of twenty-five well-documented, well-screened, mutually consistent studies can be a review of twenty-five shared errors, and the four artefacts above will look immaculate. Evaluating the sources is a separate test with separate questions, run on each document before its claim enters the synthesis.
The decision
Write the search record while you search, and accept that the first line of it will be a search you would not run again. Write the inclusion rule before the first record is screened, and put the four conditions in it that a stranger could apply. Log every exclusion with one reason and one stage, and make the funnel reconcile before the synthesis is drafted. Then write the gap statement before the conclusion, so that the conclusion sits inside the ceiling the gap statement sets rather than beside it.
The test to apply to your own review is one sentence long: hand the search record, the rules and the funnel to somebody who was not involved and ask them which literature the review cannot see. If they can answer from the four artefacts, the review is defensible at the level it claims. If they cannot, the review is a synthesis with a reference list, which is a different and much commoner thing.
The agent’s part is bounded and worth stating. It screens records against a written rule, reads the included documents into their own text, drafts the synthesis with each claim tied to a passage, and keeps the funnel’s arithmetic so that the counts reconcile. It cannot search, and it cannot decide what a missing piece of literature means. Those belong to the person who runs the searches and to the person who reads the gap statement and decides whether the review is worth commissioning further.
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 station that keeps the record: studio.x402.press/roles/researcher — the register, the claim list, the cross-check and the three refusals
- The station that builds on it: studio.x402.press/roles/curator_editor — study material, the citation matrix 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
- The argument the review feeds: how to write a whitepaper — the evidence under each claim, and the section most of them omit
- The shape of the write-up: 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 for a short prompt on deepseek-v4-flash, which is $0.00700, and 9550 under upto at max_tokens: 1000, which is $0.00955; qwen3.8-flash’s $0.00875 was read from its own challenge at the same settings. The call counts — ten records a screening call, one call a full text, six calls for the synthesis, two cross-checks — are planning assumptions constructed for this piece, not observations of a real review.
The review in Tables 2 to 4 does not exist. Its eight searches, ninety-six records, four exclusion reasons and twenty-five included sources are constructed to show the funnel reconciling, and no field, journal, standard or database is named as having been searched. No figure about publication bias, screening accuracy or inter-rater agreement appears, because none has been measured here. The four artefacts are the register and claim-list structures the research station works to, applied to a review; the limits in the closing section are the publication’s own statement of what that method cannot carry.
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: literature-review, method, sources, research
