Citation and Evidence Alignment: Connecting Every Verifiable Claim to the Minimal Sufficient Source Text

## Citation is a mapping relationship, not answer decoration

A paragraph ending with three links looks substantiated, but upon careful checking you may find that not a single sentence is actually supported. The reason is that a link can only prove that "such a document exists"; it does not tell you which sentence in that document supports the adjacent claim. The document may be about a completely different topic, may mention the term only in passing in a footnote, or may even contradict the claim. Links themselves do not carry semantic evidence; treating links as evidence is itself a confusion.

A correct citation is a verifiable mapping relationship: starting from a claim in the answer, it maps to a specific passage in the source document, and that passage semantically entails the claim. By entailment we mean that after reading the original text, a rational reader would accept that "it supports this claim," not that "it merely mentions the same word." To maintain this mapping, the system must preserve four kinds of information: document identity (which document it is), version (different versions of the same document may have different content; citations must be locked to a specific version), page or paragraph location, and the exact character span. With the character span, readers can jump back to the original sentence in one click and confirm with their own eyes that "the original text really says this."

The inputs and outputs of this mapping are clear. The inputs are three things: the decomposed atomic claims, the evidence whitelist provided by the retrieval stage, and stable document spans within that whitelist. The output is a clickable mapping from claims to versioned original text. The causal chain is one-way: first the actual retrieved document spans exist, then the support determination for the claim is made, and only then are the links rendered. The order cannot be reversed.

The model's role in this process is selection, not creation. It must not generate URLs from memory—URLs in model memory may be hallucinated, long expired, or replaced mirror sites. The correct approach is: the model selects evidence IDs from the retrieval whitelist and identifies the span within that evidence that supports the claim; link generation is left to the rendering layer, which looks up the evidence ID in a table and constructs the clickable address. This separation of "evidence selection" and "link presentation" ensures that errors in one layer cannot masquerade as success in another.

From this we can derive an explanatory rule: the mere existence of a link only proves that a document can be located; a span that semantically entails the claim counts as real support. When checking against this rule, the question to ask about each claim is not "are there nearby links," but "does the sentence the link points to really say this." When the whitelist contains no evidence that can support a claim, the system has only three acceptable actions: delete the claim, downgrade it to an uncertain statement, or refuse to answer. The fourth action—adding a topically related link to make up the numbers—is precisely the most common mistake. Topical relatedness is not semantic support; a link about "climate change" cannot support a specific figure like "rainfall in a certain region decreased by 30%." Adding a link creates the appearance of support and masks the fact that evidence is missing, which does more harm than honestly saying "no source was found."

## First split compound sentences into atomic claims

Compound sentences are the greatest enemy of citation alignment. "The new model is faster, more accurate, and costs 30% less"—how many claims need evidence in this sentence? At least three: faster, more accurate, and 30% cost reduction. These three can each be true or false independently: a model might indeed be faster but show no change in accuracy, or more accurate but double the price. Treat the whole sentence as one citation unit and you will never know which link supports which part.

The product of splitting is atomic claims: the smallest statements that retain complete meaning and can be judged true or false independently after removing connectives. Speed, accuracy, and cost change are three atomic claims. After splitting, check them one by one; each claim is bound only to the minimal sufficient evidence that supports it—if one piece of evidence is not enough, add more; if there is no evidence, delete, downgrade, or refuse to answer according to the rules in the previous section.

Which atomic claims need citations? There is a practical rule of thumb: numbers, dates, attributions, causality, and comparisons usually need citations. Numbers are the most typical—the 30% in "cost reduction of 30%" must have been measured by someone; dates involve temporal positioning of an event; attributions involve "who said it, who did it"; comparisons involve "faster than whom, faster than which version"; causality involves "because of what, therefore what". Once these types of claims are false, the error is usually falsifiable and consequential.

Connectives are the most hidden trap when splitting, because they quietly introduce additional assertions. "Because A, therefore B" appears to be one sentence but actually contains two claims: A is true (factual claim), and there is a causal relationship between A and B (causal claim). The part after "because" requires causal evidence, not evidence that "A did happen"—a document proving A occurred does not automatically prove A caused B. Similarly, "therefore" often marks the author's own inference: the first half may all have sources, but the part after "therefore" is a new conclusion the author drew on the basis of the first half, and it needs to be examined separately for whether it has additional evidence or is merely a system inference.

When splitting, also label each atomic claim with its type: statement requiring citation, system inference, or subjective suggestion. The type determines how it is handled. Subjective suggestions themselves can go without citation—"recommend adopting layered caching" is an opinion, not a fact, and forcing a citation would seem false. But the factual premises supporting the suggestion still need citations: if the reason for the suggestion is "layered caching can reduce latency by 40%", this 40% is a verifiable factual claim and must have its own evidence. At the same time, clearly distinguish in the wording which sentence is a source statement and which is a system inference: inferences can be questioned and corrected, and if they are disguised as source statements, readers lose the opportunity to question them.

The input of the whole process is the compound sentence in the answer, and the output is three things: atomic claims that can be judged true or false separately, a type label for each claim (requires citation / inference / suggestion), and a list of evidence requirements for claims that require citation. This list is the input for subsequent retrieval and matching, and its quality directly determines the upper limit of the entire citation system: implicit claims missed during splitting cannot be recovered by any later step.

## Worked Example: Calculating Citation Precision and Coverage

An answer contains four claims C1–C4 that should be cited, and the system provides three citations E1, E2, and E3. Checking one by one: E1 supports C1, E2 is only topically related to C2 but does not entail it, E3 supports C3, and C4 has no citation. How well does this system actually do? You cannot answer based merely on “three links seem like a lot”; you must calculate two metrics separately.

Lay out the four claims and three citations; the basis for each judgment is semantic entailment, not topical similarity:

E1 and E3 are truly correct citations: reading the original evidence is sufficient to infer the corresponding claims. E2 is a typical topic-related failure—a pricing homepage is certainly related to cost, but it does not contain the number 30%, nor does it contain any “before-and-after comparison” of a decrease; the reader cannot infer C2 from that page. C4 has no evidence at all.

On this basis, define two metrics. Citation precision (CitationPrecision) measures how many of the provided links truly support their claims: two of the three provided citations are correct, so precision = 2 ÷ 3 ≈ 66.7%. Coverage (Coverage) measures how many of the claims that should be cited receive correct support: among the four claims that should be cited, C1 and C3 are supported, so coverage = 2 ÷ 4 = 50%.

Written in general form: the inputs are four counts—the total number of claims that should be cited Nclaim, the total number of provided citations Ncitation, the number of correctly supported citations NsupportedCitation, and the number of correctly supported claims NcoveredClaim. The outputs are:

CitationPrecision = NsupportedCitation ÷ Ncitation; Coverage = NcoveredClaim ÷ Nclaim

The numerators of the two fractions have different meanings: the numerator of precision counts “citations” (one incorrect citation lowers it), while the numerator of coverage counts “claims” (one missed citation for a claim lowers it). The denominators also differ: precision looks at the set of links the system provides, while coverage looks at the set of claims in the answer that should be supported.

66.7% and 50% describe two different things and cannot replace each other. 66.7% describes the quality of the provided links—most of the citations given truly support their claims; 50% describes the coverage of claims that should be cited—half of the claims that should have evidence are not supported. A system can provide one high-quality citation but cover only a quarter of the claims (precision 100%, coverage 25%), or it can have a link for every claim but half are topical padding (coverage may be inflated by rough counting, while precision exposes the truth). Both metrics must be considered together: precision prevents using irrelevant links to inflate coverage, and coverage prevents using a few precise citations to mask large-scale missing evidence.

## Evidence anchors must withstand chunking, updates, and rearrangement

Many citation systems store chunk IDs directly in citations: evidence is in the 7th chunk of the retrieval corpus, so the citation records "chunk_7". This is fine at build time, but chunks are indexing artifacts, not properties of the document itself. Once the retrieval corpus is re-chunked—whether the chunking strategy changes, the window size changes, or upstream documents are updated and re-indexed—the content of the 7th chunk changes. Clicking old citations then jumps to unrelated text, or simply fails. Chunk ID is just a temporary coordinate generated by the index for retrieval convenience; treating it as a permanent citation address is mistaking scaffolding for the foundation.

Stable citations need to be anchored to the document's own original coordinates, not to index coordinates. A citation that can withstand chunking and updates should store five things: a normalized document ID (the same document points to the same identifier regardless of how it is indexed), a content version or content hash (locking in which version was seen when the citation was made), the original page number or section location, the exact character span, and a snippet (a copy of the sentence in the original that supports the claim). Chunk is only a temporary container during retrieval; the correct approach is to store a "chunk → original anchor" mapping: after chunking changes, the citation first retrieves the anchor through the mapping, then locates the original text. That way, no matter how chunks change, the citation points to the same sentence.

Web sources must also contend with updates. Online documents can be edited, rewritten, or even taken down. When crawling web pages, you must record both the crawl time and a snapshot identifier; the citation points to the snapshot, not just the URL. When the old version can no longer be located—the page has been redesigned, or the snapshot has expired—the system must explicitly flag link drift rather than silently redirecting citations to the new document homepage. Redirecting to the homepage is abandoning evidential responsibility: readers who click the link see an article that has already changed and think it is the original.

PDF documents are more complicated. In tables, footnotes, and OCR-extracted text, plain-text character offsets are not sufficient to reconstruct visual context: a table in the OCR text stream is broken into scattered numbers, and an offset like "page 3, line 2" may fall in the middle of the table, leaving readers unable to match anything when they jump there. Such sources need layout coordinates or structural paths as part of the anchor to locate the citation at a visually recognizable position. Footnotes are similar—they may be far away from the body in the text stream, but the citation needs the structured position of "which footnote".

Citation click failures are themselves signals and should become monitoring metrics. Every time a user clicks a link, it is an implicit verification; landing on the wrong location, returning to the homepage, or 404 are all direct evidence that the anchor system is broken somewhere. Tracking click failure rates and aggregating failure locations by source document can identify chunking errors, snapshot expiration, or coordinate mapping bugs before citations fail at scale. A citation system that cares only about "links looking right at generation time" and never observes "whether clicks actually reach the original sentence" will eventually have its anchors collectively drift out of alignment in some inconspicuous index rebuild.

## Original figure: Select evidence before generation, verify each claim after generation

"Write the answer first and add links later" is a breeding ground for false citations. In this order, the answer text is already fixed, and links are merely decorations pasted on afterward; the model has no incentive to check whether the evidence entails the claim, and the least-effort approach is to search by topic and pick a few documents whose keywords are tangentially related. Topically related links are misleading to readers—they all look relevant, but none of them actually says that sentence. To block this path, the causal order of citations must be reversed: evidence before answer, verification before rendering.

The generation process can be represented as a one-way path. The user question first passes through evidence retrieval to obtain an evidence whitelist; the whitelist and the question then enter atomic claim generation, producing a set of claims whose truth can be judged independently. Claims and evidence converge here to form a claim-evidence bipartite graph: one side has claim nodes, the other has evidence nodes, and an edge represents "this evidence entails this claim". After the bipartite graph is formed, first perform an entailment completeness check—whether every claim that should be cited has an edge connected to sufficiently strong evidence—while also checking source quality, and only then enter the rendering layer, turning verified edges into links with locations.

The core structure expressed in Figure 1 is: the answer is not flat text of "text + links", but an auditable claim-evidence bipartite graph. Text is only the expression layer for claims, and links are only the rendered form of verified edges in the graph. Auditing happens at the graph layer: check the entailment relationship of each edge, check for isolated claim nodes, and check the source quality of evidence. These structures are not visible at the text level, so performing citation checks only from the text is working at the wrong level of abstraction.

The whole process forms an evidence closed loop: the inputs are the question, the evidence whitelist, and atomic claims, and the outputs are a claim-evidence bipartite graph, entailment completeness check results, and located links. The loop's constraints are at both ends. Before generation, select evidence first, restricting the model's language space to available materials—the model can only write claims based on real document spans in the whitelist, and cannot produce specific numbers that are not in the evidence, because those materials are not present at generation time. After generation, verify claim by claim to catch omissions: even with generation-time constraints, the model may still write sentences not entailed by the evidence, or bind two claims to the same piece of evidence that supports only one of them. During verification, for each claim check whether its edge in the bipartite graph exists and whether the entailment is sufficient.

Failed claims have three handling methods: deletion, downgrading, or supplementary search. Deletion removes a claim without evidence from the answer; downgrading rewrites it as an uncertain statement—"according to reports", "not independently confirmed"—so that it no longer appears in a factual tone; supplementary search is for claims that lack sufficient evidence but are too important to delete, going back to search another round. Among the three methods, there is no "attach a topically related link to it"—that would be drawing a nonexistent edge in the bipartite graph, showing auditors a support relationship that does not exist.

## Entailment, Source Quality, and Factual Truth Must Be Judged Separately

A low-credibility self-media article indeed wrote the sentence "a certain drug is effective", and the model cited this sentence under the claim "the drug is effective". Is this citation correct? It depends on which layer the question belongs to. From an alignment perspective, the citation is faithful—it accurately supports the attribution statement "the article claims that the drug is effective"; but from a factual perspective, whether the drug is effective is still unknown, and a low-quality article's assertion does not constitute any factual evidence. Both judgments are true; they answer different questions. Mixing them together yields two contradictory impressions: "citation correct" and "conclusion dubious".

Correct verification is divided into three layers, each independently outputting a result. The first layer is entailment judgment: whether the specific span of evidence semantically entails the claim. It only answers "Did the original text say this?" and does not involve whether the original text is correct. The second layer is source suitability: whether this source is suitable for supporting this claim. A forum post entailing "the software has a vulnerability" does not mean it can support this claim—vulnerability statements require primary sources, original research, or official records. Source suitability also checks publication date (is a review from three years ago still valid for today's version?), sample (can a small laboratory sample be generalized to production environments), and scope of applicability. The third layer is factual truth: whether multiple mutually independent sources jointly support this fact. A single primary source may also only apply to a specific sample and a specific time; factual credibility requires cross-source corroboration.

The relationship among the three layers is progressively deeper, but they do not entail each other. The article indeed claims X, which only proves the first layer (attribution faithfulness) holds and does not prove that X is true; a primary source may pass the second layer (source suitability), but what it records may be only an observation under a specific condition, and the third layer still depends on the degree of convergence among independent sources. Conversely, multiple sources jointly supporting a fact does not mean that a particular citation is qualified at the entailment level—the citation may still be attached to the wrong span.

The three levels correspond to three different sets of inputs. The input for entailment judgment is the claim and the supporting span; the input for source suitability is the claim and source attributes (source type, publication date, sample description); the input for factual credibility is the claim and multiple independent pieces of factual evidence. The outputs are three results that can each hold independently: entailment judgment, source suitability, factual credibility. If any one of them is negative, it means the answer has a defect at that layer and requires different remediation—entailment failure requires replacing evidence or revising the claim; source unsuitability requires replacing the source or downgrading the statement; factual conflict requires returning to the conflict-handling mechanism.

Automatic entailment models can only play a screening role in this system. They are good at quickly judging "whether this text broadly supports this sentence", but they are unreliable on several types of difficult linguistic phenomena: negation words ("failed to reproduce" and "reproduce" differ by only two words), tabular data, scope qualifiers ("under specific conditions", "some samples"), and inferences that require cross-sentence composition to hold. These cases must be caught by manual spot checks. Automatic screening filters out a large number of clearly failed cases, leaving boundary cases that require human review; the two are in a pipeline relationship, and if either link tries to replace the other, it will fail in its own particular way.

## Conflicting evidence cannot be silently averaged

Two equally credible sources give different figures: A reports a market share of 18%, and B reports 24%. Which one should the model choose? Directly selecting the item that ranks first in the retrieval results is the least effortful and most common mistake. Retrieval rank measures query relevance, not evidence credibility, and even less the authority to adjudicate facts. Ranking higher only means it is easier to find; it has nothing to do with whether it is closer to the truth.

The first step in handling conflict is not to choose, but to check whether the conflict really exists. Different figures often arise because the two sides are not actually describing the same thing: different scopes (one counts installed base, the other counts active users), different time periods (one uses quarterly data, the other annual data), different versions (metrics before and after a product revision), different populations (global market versus a particular country's market). After checking scope, time, version, and population one by one, a considerable portion of "conflicts" will naturally dissolve—18% and 24% may respectively correspond to two differently defined markets, and both can be true at the same time.

If, after checking, the conflict still genuinely exists, the correct output is to present them side by side, not to adjudicate. The answer should list A's claim, its evidence, and applicable conditions, as well as B's claim, its evidence, and applicable conditions, and clearly state that the system cannot adjudicate between the two. Concealing either side is tampering with the evidence landscape: readers have the right to know that credible sources disagree. Writing "side-by-side presentation" as "the answer is longer" is laziness; this extra length is precisely the minimum cost of honesty.

When a combined inference must be made, the inference must be labeled as an inference. For example, if A and B give different figures but point to the same trend, one can write "Inferred from data from A and B, the metric shows an upward trend"—the words "inferred from A and B" must not be omitted. Once omitted, the inference masquerades as a source's original words, and readers will take the system's own conclusion as A's authoritative statement. The distinction between source statements and system inferences is once again a hard requirement here: any conclusion that goes beyond the original text of a single source must indicate from which sources and by what means it was derived.

The input to the entire conflict-handling process consists of figures from multiple sources, along with their respective scope, time, version, population, and authority information; the output has only two forms: for those that can be adjudicated after checking, give a choice with reasons; for those that remain conflicting after checking, give a side-by-side explanation that preserves both sides and their respective applicable conditions. A third form—taking an average, or picking one that looks good and writing it in—is unacceptable in terms of correctness: the average may fall at a position supported by no source, while silent selection disguises an open disagreement as established fact.

## Citation Interface Is Also Part of the Verification System

When a user clicks a citation marker and lands on the document's homepage, having to search for that sentence from the beginning—is that verifiable? No. The whole point of verification is that readers can return to the supporting span itself and compare the answer's claim side by side with the original sentence. Going only to the homepage is equivalent to shifting the work of "finding the evidence" from the system to the user, and the homepage may be a report hundreds of pages long or a page with thousands of lines; most users will give up at that step. The lower bound of verifiability is not "the link can be opened," but "after clicking, it highlights the original passage that supports this sentence."

Each part of the interface has a corresponding responsibility. The citation marker must be immediately adjacent to the claim it supports—a marker dangling at the end of a paragraph, where readers cannot tell which sentence it refers to, is a pseudo-citation manufactured at the interface layer. On hover, it should display a short excerpt, the source name, and the date, so readers can judge without clicking whether the evidence is worth opening; after clicking, it should jump to the original sentence located by the anchor and highlight its context, so the entailment relationship can be confirmed with one's own eyes. When multiple claims share the same piece of evidence, the interface must clearly mark which sentences the evidence covers; when the same claim requires multiple sources, the sources should be presented side by side, rather than providing only one "representative" link.

Mobile devices are a real constraint on citation interfaces. If citation excerpts and details are visible only in the desktop hover state, touchscreen users completely lose this layer of information. Mobile must provide expandable interactions so that summaries, sources, and jump capability are equally reachable in an environment without hover. Interface design is not an aesthetic issue here: if users on any platform cannot verify, then the citations output on that platform are equivalent to no verification in that dimension.

The presentation of excerpts is constrained by two boundaries. First, minimal sufficient context—the excerpt contains only the original text fragment sufficient for readers to judge the entailment relationship, not copying the whole paragraph; this both protects readers' attention and avoids excessive copying of copyright-protected content. Second, honesty about access permissions—for paywalled or restricted sources, the interface must state the access limitation and must not use excerpts that cannot be publicly verified to pass themselves off as open evidence. Providing a piece of "evidence" that others cannot open and implying it has been verified is to create a false exit in the verification chain.

Therefore, the input to the citation interface is not just a link address, but five things: the claim, the minimal excerpt, the source and date, the location anchor, and the access permission. The output corresponds to three interaction layers: the marker immediately adjacent to the claim, the summary visible on hover (or on mobile tap), and the location that highlights the original sentence after clicking. Only when all three layers hold can readers truly complete the round trip from answer to evidence—and the system's correctness metric is defined precisely by whether this round trip can be completed.

## Evaluation needs human-labeled claims and must also test whether the location is truly usable

Automatically checking whether each citation URL returns 200 is the entire citation evaluation of many systems. The most serious error this check can catch is "a completely dead link," which is precisely the error users can most easily discover on their own. Returning 200 only proves that the resource is accessible: the link may point to the document's homepage rather than the supporting sentence, may point to an unrelated paragraph in the same document, or may point to an article that is superficially related but says nothing. Using accessibility as a substitute for correctness is like testing only the easiest part of the verification system and letting the hardest parts pass.

Evaluation must be built on a human-annotated dataset. Annotators need to do much more than "look at links": decompose the answer into claims requiring citations, annotate the minimal sufficient evidence for each claim, judge the entailment relationship between evidence and claim, and annotate source quality. This annotated dataset is the gold standard for citation evaluation—without it, entailment cannot be objectively determined, and all metrics degenerate into formal checks.

On the annotated dataset, report six types of metrics separately: citation correctness and completeness (using the formulas from earlier), invalid or misplaced anchor rate (the proportion of citations that cannot reach a supporting span after clicking), source tier (distribution across tiers such as primary sources, original research, official records), conflict coverage (whether conflicting evidence is presented side by side when encountered), and the task quality of the answer itself—because if the task quality is very poor, no matter how standard the citation format is, it is meaningless. Metrics should also be sliced by difficulty dimensions: numbers, dates, causality, tables, and multi-hop inference each form a category. Numbers and dates are the basic difficulty of entailment judgment; causality, tables, and multi-hop inference are precisely the scenarios where automatic entailment models are most error-prone. If not sliced separately, systematic failures in these categories will be diluted by the overall metrics.

The evaluation set must include questions with "insufficient evidence." A system that refuses to answer or downgrades its wording when evidence is insufficient is part of citation discipline; if evaluation is limited to "every question must generate an answer with citations," the system will be trained to force citations for any question. The result is to improve formal coverage while producing more hidden pseudo-citations—every claim has a link, every link returns 200, but entailment relationships largely do not hold. This type of error is harder to discover than an honest refusal: it passes all formal checks and can only be caught by human entailment annotation.

The inputs to citation evaluation are therefore four human artifacts: decomposed claims, gold-standard evidence, entailment annotations, and source quality annotations; the outputs are six types of metrics and their slices by dimension. Link accessibility is only a weak approximation of the misplacement rate among them. Whether evaluation resources are spent on human entailment annotation or on expanding the coverage of automatic URL checks determines the direction in which the system is ultimately optimized: the former optimizes "citations really support," the latter optimizes "citations seem clickable."

## First clarify: retrieval, citation, attribution, and fact-checking are four different gates

The system clearly retrieved the correct paper, yet the answer may still be unreliable. Such failures are confusing because they treat four different gates as the same thing.

Retrieval does only one thing: bring candidate materials into context. It guarantees that "this material was seen by the system", nothing more. Citation connects a claim in the answer to a span of material, guaranteeing "this sentence has a source". Attribution faithfully expresses "who said what", ensuring no substitution between source and expression—after retrieving the correct paper, the system can perfectly write the experimental group's numbers as the control group's numbers; at that point the citation is real, but the attribution is already distorted. Fact-checking goes one step further, judging whether the statement itself is credible: the system can perfectly cite a paper that has been retracted, with both citation and attribution passing, yet the conclusion rests on invalid evidence.

The failure modes of the four gates are different, and the places to fix them are also different. Not finding the key document is a retrieval recall problem; the answer shows a lack of evidence. Citing a nearby number—the link is real but the original sentence refers to another metric—is a span and entailment problem. Writing the author's hypothesis as a conclusion is attribution distortion, which is a claim-type labeling error. Using only outdated secondary sources is a source and fact-checking problem: the claim has support but can still be wrong. The four gates must be separately recorded as pass or fail; a blanket "grounded" label will paint the four types of failure the same color, making it impossible to locate which layer to fix.

The distinction between evidence and interpretation is also in this set of concepts. Evidence is locatable material; interpretation is synthesis based on material, which draws on multiple pieces of evidence and adds inference. Interpretation can be very valuable, but it must explicitly mark inference steps, applicable conditions, and uncertainty. Statistical conclusions are where interpretation distortion is most concentrated: for conclusions like "no significant difference", citing only the adjacent sentence in the abstract that says "results are not significant" is far from enough—the abstract may omit the comparison object, metric, confidence interval, and sample information. The citation must support not the phrase "results are not significant", but "not significant on what comparison, what metric, and what sample". If the check scope is inconsistent with the claim scope, the citation does not hold at the entailment level.

What beginners most confuse is exactly this: "the answer has a citation" at most indicates that a link exists. Only when claim splitting, span entailment, source quality, and conflict checking all pass does the answer approach verifiability. The conceptual distinction of the four gates divides input into four parts: retrieval results, citation spans, attribution expressions, and fact-checking evidence; output is the pass or fail status of each of the four gates. If any one gate fails, it will appear in the final answer with different symptoms, and none of them can be compensated by the passing of the other gates.

## Connect the Causal Chain

Every link in the citation system depends on the output of the previous link and determines what the next link can do. Only by walking through the entire causal chain can you see exactly how cutting corners at a particular step will ultimately appear to the user.

The starting point of the chain is identifying the atomic claims in the answer that should be cited. Compound sentences are split apart, implicit assertions introduced by connectives are singled out, and each claim is labeled with a type of "needs citation, inference, or suggestion". A claim missed at this step will not be recovered in any subsequent step—retrieval does not know what evidence to look for, and verification does not know what entailment to check.

Only with claims do we have the input for retrieval: find candidate evidence from the whitelist. The whitelist is the boundary of this step; the model can only select from it and cannot generate URLs from memory. The candidate materials brought back by retrieval carry document identity and are then solidified into stable citation anchors—document version, page number or section, character span, and minimal excerpt are kept together so that citations do not depend on chunk IDs that can change at any time.

Only after anchoring is fixed does entailment judgment come: does the specific span of evidence semantically support the claim. The basis for judgment is whether the span content can entail the claim, not whether the topics are similar. Evidence that passes the entailment check must also pass the source quality gate—publication date, sample, scope of applicability, source tier, and conflicts with other sources. Conflicts cannot be silently averaged; when they cannot be adjudicated, present both sides and their applicable conditions side by side.

After all verification is complete, there are only three ways out when evidence is missing: delete the claim, downgrade it to an uncertain statement, or refuse to answer. Adding a topically related link is not among these three—it disguises verification failure as verification success. Only edges that pass verification enter rendering: citation markers sit next to claims, hovering shows the excerpt and source, clicking locates and highlights the original sentence, and mobile devices can expand it just the same. Rendering is the exit of the causal chain, but it can only present the support relationships that have already been established upstream.

The chain does not end at rendering. Sample-checking citation correctness and completeness operates the whole chain as an observable system: how many of the given citations truly support, how many of the claims that should be cited are covered, and whether the click failure rate is rising. Problems found in the review are traced back along the causal chain—misalignment rate points to the anchor layer, entailment failure points to the verification layer, and low coverage points to the splitting or retrieval layer. The output quality of each link in the chain is ultimately measured by this closed-loop review, and the measurement results in turn become the input for the next round of correction. The whole system is therefore not an action of "pasting links at generation time" but a continuous round-trip path from claims to evidence and back to claims.