20.1The claim being checked
Checking begins with the sentence the author wants supported, but that sentence is rarely one claim. Separate its small claims first: what the cited passage actually reports, what the author draws from it, and what further proposition would be needed to carry the author's wider wording.
Take the invented report R1, section 2, which records that of ten respondents, six requested later hours. A candidate may report that limited result. The same evidence does not license describing 6 of all service users; that step requires a population premise, such as a claim that the ten respondents represent the service-user population. If the sentence further says the clinic should extend hours, that is a further inference again, and R1 alone does not supply its warrant. None of these steps is established merely by having a link to R1: the link locates the excerpt, and the judgment that the excerpt entails each smaller claim is a separate act by the reader.
The same separation settles what looks like contrary evidence. The invented report R2 concerns a different year's respondents. Its proportion need not contradict R1's, because the two passages concern different respondent sets; calling them contradictory would itself require a premise connecting those sets or a shared measure and population. No external truth audit is performed here on either invented report. The question is only whether a cited passage, taken as given, supports the claim placed beside it, and what that judgment depends on.
Once the small claims are separated, each can be matched to a located passage, and gaps become visible as missing premises rather than vague inadequacy. That visibility matters because the author's defensible inference may still be sound even when one premise has no source that states it outright. But decomposition is diagnostic rather than final: the clauses' relations must be recombined and the resulting inference judged as a whole.
Once the sentence has been separated into small claims, the remaining risk sits in the joints. The clauses of a claim carry relations—about population, period, comparison, and consequence—that no single fragment states on its own. Decomposition can therefore make each fragment look adequately supported while the sentence assembled from them says more than the evidence allows.
Take report R1, section 2, which reports that of 10 respondents, 6 requested later hours. A candidate claim that faithfully repeats that result survives the first check. But a claim that 6 of all service users requested later hours requires a population premise linking the 10 respondents to the larger group. Without that premise, the step from the located passage to the broader claim is not an entailment; it is an inference, and an unsupported one as stated. The reader should be able to see which part is the passage's result and which part is the author's extension.
Relations also govern disagreement. Report R2 concerns a different year's respondents, so a proportion that differs from R1's does not contradict it: differing numbers become contrary evidence only if population and period are held fixed. Neither invented report is audited here for external truth; the check concerns what the located passage licenses.
The corrective is modest. Atomic checking remains useful for finding which premise a locator supports and for exposing an empty assertion. What it cannot settle is whether the premises, once recombined, warrant the conclusion attached to them. So the recombined inference deserves its own check, with three possible verdicts: entailed by the cited passage, defensible as the author's inference given stated premises, or not yet warranted because a population premise is missing. Not every original analytical sentence needs a source that already says it, but the absence of such a source makes the judgment of defensibility more, not less, important to state.
20.2Finding the relevant passage
Once an argument has been decomposed into candidate claims, the item still requiring judgment is the inference those claims jointly support. Consider the invented report R1, section 2, which reports that of 10 respondents, 6 requested later hours. The recombined claim to be checked is that later hours were requested, with the stated basis of six requests among ten respondents. A locator's first task is to place an authorized reader at that sentence in that section, not merely somewhere inside R1. The reader must be able to tell which words supply the proportion, the denominator and the measured response.
The located passage supports that limited result. It does not by itself support the broader statement that the same six-in-ten proportion holds among all service users, because that extension requires a population premise connecting the respondents to service users generally. Here the locator and the entailment judgment divide the work: a locator can show where a population premise appears, if it appears at all, but it cannot manufacture the premise or decide that the inference from respondents to all users is valid. Without an additional justified population premise, citing this passage alone does not support the wider claim.
The same separation applies to conflicting material. Invented report R2 concerns respondents in a different year, so its differing proportion need not contradict R1's. A locator enables the comparison by exposing each report's population and period; whether two observations actually conflict is a further judgment, and a mismatched year or group is not contradictory evidence merely because the numbers differ.
ALCE separates dimensions of cited-answer evaluation and checks support rather than merely counting citations (Gao et al., 2023). That work does not certify every reference, nor does adding citations guarantee correctness. Its useful lesson for locating evidence is narrower: support should be examined at the passage level, where the premise actually used by an inference can be identified and checked.
A correct title or a functioning URL answers only the first question: which document is at issue? It does not yet identify the passage on which the claim rests. Consider invented report R1, section 2, which states that of ten respondents, six requested later hours. A citation to R1 by title or URL may be entirely accurate and still require the reader to search the whole report before finding that sentence. The located passage, not the document, is the unit that lets an authorized reader check the stated premise.
Scope follows from the passage that is actually located. Section 2 licenses a report about ten respondents and six requests for later hours. It does not license the broader description that six of all service users wanted later hours; that extension needs an additional population premise, which the citation does not supply. A reader who cannot see which passage was relied on cannot tell whether the author kept to the local result or silently enlarged it.
The same discipline governs comparisons across reports. Invented report R2 concerns respondents from a different year, so differing proportions in R1 and R2 need not contradict one another. That question belongs to the assessment of support and contradiction; what matters here is that each citation must fix the premises being compared before such an assessment can begin. Nothing in this locating work audits the truth of either invented report, and no external truth audit is performed.
Finally, the passage is a premise, not the whole reasoning built upon it. A citation to one premise does not own the inference the author draws from it. The inference remains the author's step, to be assessed on its own terms and stated with any assumption it requires.
20.3Support and contradiction
Once the passage is located, the next judgment is not about the document as a whole but about one specified claim. Take report R1, section 2, which records that six of ten respondents requested later hours. That excerpt supports the narrow claim that, among those ten respondents, six asked for later hours. It does not support the claim that six of all service users wanted later hours; that extension needs a population premise linking the ten respondents to service users generally. A citation to the located excerpt therefore fixes the evidential reach of the claim the reader may check, while leaving the generalization as the author's inference.
The same discipline separates contradiction from silence and from mere difference. R2 concerns respondents in a different year. If its proportion is lower or higher, the two reports need not contradict: they describe different samples, and nothing in the stipulated record makes one answer the other's claim. To call them contradictory, the claim and the cited passage must bear on the same question under conditions that make their outcomes incompatible. Silence is a third status: an excerpt that says nothing about the specified claim leaves that claim uncorroborated by this citation, not refuted by it. No external truth audit is performed on either invented report here; the point is what a reader can judge from the located excerpts, not which survey is correct.
This distinction also answers the worry that original analysis lacks a source that already states it. An author may defensibly infer, for example, that later hours respond to the reported demand, provided the inference is labeled as such and its premises are citable. The link, the passage, and the entailment judgment remain separate; only the last two support the analytical sentence.
Scope determines what a located passage can carry. In invented report R1, section 2 states only that of 10 respondents, 6 requested later hours. A note may report that result accurately. It may not turn the result into a claim that 6 of all service users requested later hours: that extension requires a population premise connecting the respondents to the wider group. Nothing in the passage supplies that premise, so topical relatedness is not support for the broader claim.
Invented report R2 concerns respondents from a different year. Even if its proportion differs from R1's, the figures do not contradict one another. Contradiction requires claims about the same population and period that cannot both hold; a difference across two samples leaves each report within its own scope. The note also contains one genuinely contradictory observation, but it earns that description only because it is directed at the same claim the candidate makes, under conditions that make the two incompatible. No external truth audit has been performed on R1 or R2, so the classification concerns how the stated claims relate, not whether either report is factually correct.
Once the candidate's claim is narrowed to what R1 actually reports — 6 of R1's 10 respondents — R1 supports it and R2 is silent. Silence supplies no verdict either way: it neither confirms the claim nor refutes it. The same applies to the population claim, which no located passage reports; its defect is a missing premise, not evidence against it.
The relevant limit is therefore one of scope. If a later passage reported an incompatible finding for the same respondents and period, the relation would become contradiction. Conditional on the passage having those conditions, it could count as contrary evidence; a differing proportion from another sample cannot. A citation that lets a reader check the passage also lets the reader check the boundary at which this judgment stops.
20.4An inference owned by the author
A citation can mark several premises while the sentence that joins them advances a conclusion that no cited work states. That is ordinary scholarly synthesis, and it need not be a defect. Consider a hypothetical case in which one source reports that readers verify references selectively and another reports that failures concentrate in long documents. An author might hypothesize that a long reference list is a poor auditing instrument when locators are absent and checking time is limited. Those additional conditions must be stated; the two findings alone do not entail the conclusion. Neither source need contain that sentence; the inference is the author’s contribution, and its warrant is the author’s responsibility.
The citations in such a sentence identify what they support, namely the premises. A reader can then check whether each premise is accurately located and whether the step from premises to conclusion is licensed. ALCE, for example, separates multiple dimensions of cited-answer evaluation and examines whether references support the claims made, rather than treating citation counts as sufficient (Gao et al., 2023). This study's point is not that this benchmark settles how inference should be judged, nor that it certifies particular references. It illustrates, rather, that support is a relation to be assessed rather than a quantity to be tallied. The inferential step in the sentence above remains the author's to defend.
Citation laundering occurs when a nearby link is asked to do the work of an absent argument. If the sentence above were followed by a citation to the study on long documents, a reader might suppose that study had established the conclusion about auditing instruments. The citation would then conceal the gap rather than expose it. Defensible inference looks different: the cited claims are individually locatable, the author signals that the conclusion is drawn from them, and a reader who rejects the step can say precisely which premise or which linkage is at fault.
Not every original analytical sentence should therefore be forced to cite a source that already says the same thing. Compression, abstraction and synthesis are precisely where an author earns a place in the discussion. What is required is not coverage but honesty of attribution: sources for the claims that are theirs, and ownership of the steps that are the author's.
The practical test appears in the work of extending a located result. Suppose the research note's R1, section 2, says: Of 10 respondents, 6 requested later hours. That excerpt is enough for a report of what those ten respondents requested. Moving from it to a claim about service users generally introduces a population premise that R1 does not supply. The author may propose such a premise, but should mark it as a further assumption and give a reason for accepting it, so a reader who rejects the broader claim can reject the assumption rather than charge the cited source with saying something it never said. Naming the assumption in the sentence keeps the citation's jurisdiction narrow and the conclusion's ownership visible.
This matters especially when the inference is built from several pieces. If one citation is attached to R1 and the conclusion also relies on a claim about the relation between selective checking and long documents, that citation covers the located premise only. The hypothetical two-study inference from the previous subsection remains an inference owned by the author; its conditions are assumptions, not findings of the studies. A reader should be able to accept the excerpt, deny the extra condition, and stop there.
The same discipline applies to apparent tension between reports. R2 concerns respondents from a different year, so differing proportions do not by themselves establish a contradiction. The reports describe different sets of respondents. Even comparable samples from a stable population can yield different proportions; a stability assumption alone would not make the reports logically contradictory. A conflict claim must identify the incompatible propositions, while a claim about statistical disagreement requires its own model and uncertainty assessment. It also marks the limit: no external truth audit of R1 or R2 is performed here; the argument concerns what follows from the excerpts and the assumptions supplied along with them.
20.5Several sources in tension
The inference noted in the preceding section belongs to its author: when a manuscript infers selective checking from observations about long documents, it should state the assumption linking those premises, so that a reader can reject the inference without accusing either source of making it. A citation to one premise does not own the whole inference. The same discipline applies when two sources seem to disagree, because apparent inconsistency often begins as a mismatch of scope rather than a collision of findings.
Consider invented report R1, section 2, which records that six of its ten respondents requested later hours. A candidate may report exactly that limited result. The candidate may not describe six of all service users, or six of any wider public, without an explicit population premise linking these ten respondents to that larger group. Now consider invented report R2, compiled from respondents surveyed in a different year. If its proportions differ from R1's, the two reports need not contradict each other at all: they may concern different populations, or measure different outcomes under different conditions. Comparing the numbers before comparing the conditions manufactures a conflict where none has yet been established.
The disciplined sequence is therefore to separate several things a citation might appear to provide. A link locates a source; a located passage fixes what that source reports; an entailment judgment says whether the passage supports the sentence attached to it; contradictory evidence exists only when genuinely matched conditions yield incompatible results; and a defensible inference adds an assumption the author must own. No external truth audit has been performed on either invented report, so nothing here decides which is accurate. The available move is narrower: determine whether the populations, periods, and outcomes match closely enough to make comparison meaningful. If they do not, the apparent tension dissolves into a scope distinction. If they do match, the conflict is real and remains live.
Once scope questions have separated the cases, the remaining task is to decide whether a conflict has survived them. R1, section 2, reports that of 10 respondents, 6 requested later hours. That passage licenses the limited result as stated; it does not license the claim that 6 of all service users want later hours without a premise connecting the sample to that wider population. R2 concerns respondents from a different year. The two proportions may therefore differ without the reports contradicting one another.
The decisive test is whether the compared findings meet the same conditions. If the population definition, period, outcome, elicitation task, and measurement genuinely match, then conflicting results count as contradictory evidence. The research note should preserve that conflict as unresolved rather than smoothing it into invented consensus. Its located excerpts remain checkable passages, the author's judgment that the findings clash is a separate entailment judgment, and any further conclusion from the pair is a defensible inference—one that must register the observation cutting against it.
The limit is equally specific: whenever one of those conditions differs, the apparent inconsistency returns to the domain of scope, and no contradiction has yet been established. Nothing here audits the external truth of the invented reports. This preserves the qualification already carried in 20.4.1, where the explicitly hypothetical two-study inference about selective checking and long documents rests on the author's assumptions: a citation to one premise does not own the whole inference.
A reader should thus be able to check each link, locate each passage, see what it entails, and judge separately whether contrary evidence has been acknowledged or an inference has outrun its premises.
20.6Citation completeness and accuracy
Citation presence, claim coverage, and evidential support must be tested separately. A reference attached to a sentence establishes only the first: that some source has been named. It does not show that every material part of the sentence is covered, nor that the cited passage supports the interpretation placed upon it.
Consider invented report R1, section 2, which reports that of 10 respondents, 6 requested later hours. A writer may accurately report that result with a citation. The same citation cannot support the broader statement that six of all service users requested later hours. That enlargement requires a population premise linking the ten respondents to the wider population. The passage supplies the result; the absent premise leaves the generalization unsupported. Coverage therefore fails even though a citation is present.
Support is a further judgment. Suppose invented report R2 records a different proportion among respondents surveyed in a different year. The reports need not contradict one another merely because their proportions differ, since the conditions are not matched. Recording both may leave the conflict unresolved; treating either as refuting the other would claim more than the passages establish. Nor does this analysis perform an external truth audit of either invented report.
The author’s inference deserves the same separation. Preserve the hypothetical two-study argument from 20.4.1 concerning selective checking and long documents: its conditions are assumptions, and citing one premise does not own the inference drawn from the premises together. A defensible inference may warrant an original analytical sentence even when no source already states that sentence. The relevant limit is that the inference must remain recognizably responsive to its cited premises.
Thus a citation can be present, a claim only partly covered, and the relation between passage and claim still contested. Each property must be checked on its own.
A response can attach a citation to every sentence and still misdescribe what its sources establish. Consider the invented report R1, section 2: of 10 respondents, 6 requested later hours. Citing that locator while describing six of all service users reports more than the passage contains. The wider claim requires a population premise, and neither the citation nor the locator supplies one. The reference is present and correctly placed, yet the attribution is wrong because it enlarges the source's claim.
Three things must be kept apart. A link directs a reader somewhere. A located passage is the text the link reaches. An entailment judgment holds that the passage supports, or fails to support, a particular statement. The first two belong to the source; the third is the author's work, and presenting it as a reported finding is precisely the misattribution at issue. This is why the objection to universal sourcing has force: not every original analytical sentence has, or needs, a source that already says it. The obligation is to mark the inferential step as the author's, not to manufacture a source for it. In the explicitly hypothetical two-study inference preserved at 20.4.1, a citation to one premise does not own the whole step; its conditions remain assumptions.
Contradictory evidence demands the same discipline. The invented report R2 concerns respondents from a different year, and differing proportions across the two reports need not contradict one another unless population, period and measure are aligned. The contradictory observation in the research note belongs beside the author's inference as evidence to be weighed, not as a conclusion already settled.
Citation completeness is therefore necessary but not sufficient. Accuracy turns on whether each statement remains no broader than the passage cited and each inferential step is labeled as the author's own.
20.7A reader-facing evidence path
A compact evidence path keeps three elements visibly separate: the claim being made, the locator for the passage relied on, and the author's inferential step. The first says what is asserted, the second says where to look, and the third marks what the author added. A link supplies none of these distinctions on its own, and a located passage is not yet an entailment judgment.
Take the invented report R1, section 2: Of 10 respondents, 6 requested later hours. A candidate may report that limited result and give its locator. The move from six respondents to the whole service population is not in R1; it needs a population premise, and the path should label that premise as the author's. Without such a label, a precise locator makes the inference look source-backed, because the reader sees where the evidence sits but not where the evidence stops. The same discipline applies to entailment: the judgment that R1 supports the limited claim is a judgment the author makes about the excerpt, not a fact stored inside it.
The path must also give contradictory evidence its own place. Invented report R2 concerns a different year's respondents, so its proportion may differ from R1's without any contradiction arising; whether the two reports bear on one claim is a further question the author answers explicitly rather than leaves to the juxtaposition. No external truth audit is performed here on either report. The path records what each located excerpt establishes and where reasoning begins.
The gain is therefore modest but real. The reader can inspect each joint in turn and see that a citation to one premise does not own the whole inference.
The compact path earns its place only where its parts can actually be reached and read. Suppose the path points to invented report R1, section 2, where the located excerpt states that of ten respondents, six requested later hours. A reader can then do three distinct things: reach the passage, compare the candidate’s wording with it, and judge whether the stated inference follows. The first is access, the second is fidelity to a located passage, the third is an entailment judgment. None of them establishes that the reported result is true.
That distinction sets the path’s scope. A candidate may report the limited result from R1, but may not describe six of all service users unless a population premise is supplied. The path does not fill that premise in; it makes its absence visible. In this way the device answers the chapter’s question by letting a reader check what is actually present: the claim as stated, the locator, the passage it depends on, and the author’s inferential step.
Access and intelligibility are therefore conditions of the path’s usefulness, not incidental virtues. If the linked material cannot be opened, or the excerpt cannot be understood without an unavailable background, the reader’s inspection stops early. Additional suppositions about what the report might have meant must be marked as conditional rather than folded into the locator.
The same care applies to adverse material. The fictional research note also contains one contradictory observation, which should remain visible beside the inference rather than being silently absorbed into it. Invented report R2 concerns a different year’s respondents, so its proportions need not contradict R1’s merely because they differ; a contradiction would require an aligned population, measure, or condition. No external truth audit is performed on either report.
The payoff is inspectability, not certification: a reader can call an inference defensible when its premises and step are exposed, while withholding assent from the claim itself.