Trace a claim to its source

Citation Telephone

Claims change as they pass from paper to paper. Pick a cited sentence, and Citation Telephone follows it back through the papers it cites, reads each one, and shows where the claim came from and what changed on the way.

Each paper is read in full when open access, and its passage is highlighted
Every hop says whether the citing sentence matches its source, goes further or changes it
Trails end at an origin, a paper that doesn't say it, or one that's closed

Any open-access paper in Europe PMC

Examples

Example trace

Traced to an origin two hops back: Ma et al., 2022.

At one hop, the citing sentence changes what its source says.

Papers read: 2; on this trail
Origins: 1; report it as their own result
Deepest hop: 2; from the claim
Changed on the way: 1; goes further or changed
Closed: 0; abstract only or not found

The trail

2 papers · stored example

The claim

“More importantly, nuclear PD-L1 can promote cell proliferation in an immune-dependent manner without interacting with PD-1. [9]”

Zeng Z et al. · 2024 · Advancing Cancer Immunotherapy through Engineering New PD-L1 Degraders: A Comprehensive Study from Small Molecules to PD-L1-Specific Peptide-Drug Conjugates

cited as [9]Changed
1. cited as [36]

2 hops from the claim·2022·Cancer Lett

Nuclear PD-L1 promotes cell cycle progression of BRAF-mutated colorectal cancer by inhibiting THRAP3

Ma R et al.

OriginEurope PMC PMID 34923044 DOI

This paper's abstract reports its own result on what it's cited for. The trail starts here.

The citing sentence

nPD‐L1 can also serve as a key nuclear protein that promotes immune‐independent cell proliferation without binding to PD‐1. [36]

What this paper says · Abstract

Colorectal cancers (CRCs) with the BRAF V600E mutation exhibit upregulation of programmed death ligand 1 (PD-L1) but fail to respond to immunotherapy targeting programmed cell death protein 1 (PD-1)/PD-L1. Recent studies have explored the intracellular functions of PD-L1. Here, we demonstrate that PD-L1 was highly expressed in both the cytoplasm and nucleus of BRAF-mutated CRC tumor cells and tissues. Nuclear PD-L1 (nPD-L1) promoted the growth of tumor cells both in vitro and in vivo. Mechanistic investigations revealed that PD-L1 translocation into the nucleus was facilitated by the binding of p-ERK. Further, nPD-L1 upregulated the expression of cell cycle regulator BUB1 via interactions with thyroid hormone receptor-associated protein 3 (THRAP3), thereby accelerating cell cycle progression and promoting cell proliferation. Moreover, BRAF V600E-mutated CRC cells exhibited upregulation of PD-L1 expression via the transcription factor LEF-1. These findings reveal a novel role of nPD-L1, which promotes cell cycle progression in an immune-independent manner in BRAF V600E-mutated CRC.Our study provides novel insight into the mechanisms underlying BRAF V600E-mutated CRC progression.

Against the claim you started with:Changed

How “More importantly, nuclear PD-L1 can promote cell proliferation in an immune-dependent manner without interacting with PD-1. [9]” compares with this paper's passage, 2 hops down the trail.

Readings behind this result

Matches34%

Partly matches39%

Goes further25%

More cautious1%

Changed1%

Not there0%

States what it's cited for: 92% · Own result: 96% · Read from the abstract. These are the model's probabilities; the rules that turn them into a result are under How it works.

Fix it with K-Dense Web

K-Dense Web, our AI co-scientist, can find a source that supports the claim as written, read the papers that were closed here, or rewrite the sentence to match its sources. Pick what to send.

Find a source that supports the claim as writtenRewrite the sentence to match its sourcesCheck the rest of the starting paper's citationsCheck the claim against the wider literature

Copy brief and open K-Dense Web 4 of 4 selected

The brief lists the claim, each paper on the trail with its identifiers and result, and the passage each result rests on. It's built from fixed templates.

1
Find the cited paper
Each reference is matched to its Europe PMC record through its PubMed ID, PMC ID or DOI, and otherwise through its title, strictly.
2
Read it
Open-access papers are read in full, tables included, with every in-text citation linked to the reference list. For closed papers, only the abstract can be read.
3
Judge the passage
An AI model picks the paragraph and then the sentence closest to what the paper is cited for, choosing only from the paper's own text, and compares it with the citing sentence.
4
Follow or stop
A paper that reports the finding as its own result is an origin. One that cites earlier work for it is followed into those references, until every branch ends.

What it is, and isn't

A check of whether each cited paper says what it is cited for, and of where a claim entered the literature.
Not a verdict on whether the claim is true. An origin can itself be wrong; a claim can be true and still badly cited.
K-Dense doesn't store or log your claims. The claim is sent with the papers' public text to be read, and the trail lives in your tab.

Method v0.2 · October 5, 2026

How a trace is made

Everything that decides a trace is published here: every question exactly as it is asked, every threshold and every rule.

The questions asked of every cited paper

Each paper is read against the sentence that cites it, given as the claim along with how it cites this paper ([14], Smith et al., 2010). Every question is about what that sentence cites this paper for, so a sentence that also covers other points is read only for its part. The model returns a probability for every option; it doesn't write text.

Long papers are split into windows of up to 60,000 characters, each asked the same paper-level questions; the strongest window wins. Paragraphs are prefixed with an id and their section: P012| [Results] ….

1. Stated anywhere (yes or no, per window)

“The claim is a sentence from another paper that cites this paper. Does this paper state what the claim cites this paper for, as its own result or as background?”

yesSome passage states what the claim cites this paper for, or part of it, even if its numbers, certainty or scope differ
noNo passage states what the claim cites this paper for; at most the paper covers a related topic

2. Closest paragraph (one of the paper's paragraphs)

“Which paragraph comes closest to stating what the claim cites this paper for?”

The options are the ids of the paper's own paragraphs, so the passage shown is always text the paper contains.

3. Own result anywhere (yes or no, per window)

“Does this paper report its own result on what the claim cites this paper for, from its own experiments, data, analysis or pooled analysis?”

yesThe paper's own study produced the result the claim cites it for
noThe paper only mentions it as background, cites it from other work, or does not address it

4. Own-result paragraph (one of the paper's paragraphs)

“Which paragraph reports this paper's own result on what the claim cites this paper for?”

The options are the ids of the paper's own paragraphs, so the passage shown is always text the paper contains.

5. Sentences (one sentence of the top paragraphs)

“Which sentence comes closest to stating what the claim cites this paper for?”

Also asked as: "Which sentence reports this paper's own result on what the claim cites this paper for?" The two answers are the candidates judged next.

6. Support (yes or no, per candidate sentence)

“Does the passage state what the claim cites this paper for, or part of it?”

yesThe passage states what the claim cites this paper for, or part of it, possibly with different numbers, certainty or scope
noThe passage is about a related topic but does not state what the claim cites this paper for

7. Comparison (one option, per candidate sentence)

“Compare what the claim cites this paper for with what the passage says. Ignore parts of the claim about the citing paper's own work or cited to other papers.”

sameThe passage backs what the claim cites the paper for: the same finding, direction, size, population and certainty For example: Different wording, detail, emphasis or rounding; Parts of the claim about other cited papers or the citing paper's own work, which don't count against the passage.
partialThe passage backs some of what the claim cites the paper for, but not all of it
strongerThe claim states more than the passage For example: More certain than the passage; Causal where the passage reports an association; A larger number or effect; A broader population, or a general rule from a specific case.
weakerThe claim states less than the passage: more hedged, a smaller number or effect, or a narrower scope
differentThe passage addresses the same question but says something else For example: A different number that is not simply larger or smaller; A different outcome, population, species, cell type or gene; The opposite direction; A finding presented with a different meaning.
unrelatedThe passage does not address what the claim cites the paper for

8. Own result (yes or no, per candidate sentence)

“Does the sentence report a result of this paper's own study, rather than describing earlier work?”

yesThe sentence reports what this paper's own experiments, data or analysis found
noThe sentence describes earlier work, background knowledge or a general statement

Asked of the passage alone, without the claim in view, which reads more reliably.

9. References to follow (yes or no, per reference the sentence cites)

“Is this reference cited in the passage as a source for what the claim cites this paper for?”

yesThe passage cites this reference for what the claim cites this paper for
noThe passage cites this reference for something else

Each reference is named in the question with its label and citation.

10. Against the claim you started with (one option, further down the trail)

“Compare the claim with what the passage says. Ignore parts of the claim about the citing paper's own work.”

sameThe passage backs what the claim says: the same finding, direction, size, population and certainty For example: Different wording, detail, emphasis or rounding.
partialThe claim makes two or more distinct points, and the passage clearly backs some of them but not all
strongerThe claim states more than the passage For example: More certain than the passage; Causal where the passage reports an association; A larger number or effect; A broader population, or a general rule from a specific case.
weakerThe claim states less than the passage: more hedged, a smaller number or effect, or a narrower scope
differentThe passage addresses the same question but says something else For example: A different number that is not simply larger or smaller; A different outcome, population, species, cell type or gene; The opposite direction; A finding presented with a different meaning.
unrelatedThe passage does not address the claim

From readings to results

A passage addresses what the paper is cited for when its support reading is at least 40%, or when its comparison is confidently one of matches, goes further, more cautious, changed (at least 60%) with “not there” below 15%.
Of the two candidate sentences, the paper's own result is used when it addresses the claim; otherwise the one with the stronger support.
A comparison is shown on the trail when its top option reaches 40%. “Goes further” and “changed” are warnings, shown only at 50% or more; below that, the passage gets the best of its other readings.

Origin: Two sentences are judged per paper: the one closest to the claim and the paper's own result on it. The paper is an origin when the passage chosen for it states what the paper is cited for (support reading at or above 0.4, or a comparison other than "partly matches" or "not there" at or above 0.6 with "not there" below 0.15) and that passage reports the paper's own result (own-result reading at or above 0.5).
Passes it on: The paper states the claim as background and cites other papers for it. The trace follows up to 3 references: those the sentence cites at or above the follow threshold, or its most likely one. When the sentence cites nothing itself, citations in the sentences just before and after it are followed if they reach the threshold.
No source given: The paper states the claim as background without citing anything for it, in that sentence or the ones around it.
Doesn't say it: The paper's full text was read and no passage chosen for the claim addresses it.
Abstract only: The full text isn't open access, so only the abstract was read. A claim missing from an abstract may still be in the paper, so this never counts as "doesn't say it". An abstract that reports the finding as its own result still counts as an origin.
No text available: Europe PMC has the record but no abstract or full text to read.
Not in Europe PMC: The reference couldn't be matched to a Europe PMC record, for example a book, a website or a paper outside the life sciences.
Already on the trail: The paper was already read on this trace; its result is shown once.
Trail stopped: The trace stopped here: 6 hops from the claim, or 24 papers read.

Limits

Full text only for papers in the Europe PMC open-access subset. Everything else is read from its abstract, which can't show where a paper got a claim.
Up to 3 references followed per sentence, 6 hops and 24 papers per trace. You can trace any other reference by hand.
Figures aren't read, and claims made only in a figure can be missed. Tables and figure legends are read.
Books, reports, websites and papers outside Europe PMC can't be read.
The model can misread, especially subtle changes such as a swapped term. Every result opens the passage it rests on so you can check it.
The method is versioned. Compare traces only within the same version.

Citation Telephone

Traced to an origin two hops back: Ma et al., 2022.

The trail

Fix it with K-Dense Web

Find the cited paper

Read it

Judge the passage

Follow or stop

What it is, and isn't

How a trace is made

The questions asked of every cited paper

From readings to results

Limits