SCIENTIFIC CLAIMS · RESEARCH CLAIMS · EVIDENCE NETWORKS · EVIDENCE GRAPHS · NANOPUBLICATIONS · PROVENANCE · CLAIM VERIFICATION · SUPPORTING EVIDENCE · CONFLICTING EVIDENCE · AI SCIENTIFIC REASONING

How Scientific Claims and Evidence Networks Work

Scientific claims, research claims, evidence networks, evidence graphs, nanopublications, claim verification, supporting evidence, conflicting evidence and provenance are becoming central ideas in modern research because a paper is no longer the smallest useful unit of scientific reasoning. A journal article may contain dozens of claims, each supported by different experiments, datasets, statistical analyses, citations and methodological assumptions. If research systems treat the entire document as one undivided object, they can tell us that one paper cites another, but they struggle to tell us which claim is being supported, which evidence contradicts it, which version is current, how the evidence was generated and whether an AI system has actually verified the claim or merely found a semantically similar paragraph.

A modern evidence network connects a scientific claim to the observations, experiments, datasets, methods, publications, evidence passages and provenance records that bear on it. Nanopublications show one formal route: package a small machine-readable assertion together with provenance describing how that assertion came to be and publication information describing the nanopublication itself. The W3C PROV family gives a general language for provenance. CiTO gives typed meanings to citations such as supporting, extending, using methods or disputing. Newer evidence-graph systems go further by representing support and challenge relations explicitly. The objective is not to turn uncertain science into rigid triples. It is to make the chain from claim to evidence inspectable, reusable and correctable.

In 2026 this problem is especially important for AI scientific reasoning, scientific fact-checking, evidence-grounded claim verification, retrieval-augmented generation and claim-level knowledge graphs. Systems such as SciFact established the idea that scientific claims can be paired with evidence sentences labelled as supporting or contradicting. Newer systems such as SciTrue emphasise source-level accountability and evidence traceability because fluent AI answers can contain unverifiable attributions or hallucinated citations. At the infrastructure layer, Crossref’s Research Nexus argues for richer relationships among research objects, people, organisations and actions; current nanopublication work extends provenance to multi-source claims; and evidence ontologies now model evidence supporting the correctness of findings. The direction is clear: scientific reasoning is moving from document retrieval toward explicit claim–evidence–provenance networks.

A claim becomes scientifically useful when another person can ask: What exactly is being asserted? Which evidence bears on it? How was that evidence produced? What conflicts with it? Which version is current? And can I return to the source?

50-second router: where to enter this article

• I need the basic model: Sections 1–14.
• I need nanopublications and provenance: Sections 15–55.
• I need support, contradiction and evidence semantics: Sections 56–70.
• I need scientific claim verification: Sections 71–100.
• I build AI/RAG/GraphRAG systems: Sections 101–132.
• I work in publishing, systematic reviews or research integrity: Sections 133–162.
• I want implementation and evaluation guidance: Sections 163–220.
• I want the advanced design layer: Sections 221–265.

The short answer

SCIENTIFIC QUESTION
  → CLAIM
  → CLAIM SCOPE + QUALIFIERS
  → SOURCE / STUDY
  → METHOD
  → DATA / OBSERVATION
  → RESULT
  → EVIDENCE SPAN
  → SUPPORT / CHALLENGE / NEUTRAL RELATION
  → PROVENANCE
  → VERSION + PUBLICATION STATE
  → OTHER SUPPORTING SOURCES
  → CONFLICTING SOURCES
  → WEIGHT + UNCERTAINTY
  → HUMAN / MACHINE VERIFICATION
  → CURRENT CLAIM STATE
  → CORRECTION / UPDATE / RETRACTION
  → AUDITABLE RETURN PATH

Part I — What a scientific claim actually is

1. A claim is not the same thing as a sentence

A sentence is a unit of language. A claim is a proposition that can, in principle, be supported, challenged, qualified or left unresolved by evidence. One sentence can contain several claims, and one claim can be expressed across several sentences. “Treatment A reduced blood pressure by 12 mmHg in adults with condition B over twelve weeks” contains at least a population, intervention, outcome, magnitude and time interval. If the sentence continues “and was therefore superior for long-term cardiovascular prevention,” a second inferential claim has been added. Claim modelling begins by separating the propositions that require different evidence.

This distinction matters for AI extraction. A language model can split prose into grammatical sentences perfectly and still produce poor scientific claims if it misses scope, qualifiers or inferential transitions. Claim identification is semantic decomposition, not punctuation processing.

2. A claim is not the same thing as a hypothesis

A hypothesis is a proposition proposed for testing. A claim is broader: it can describe an observed result, a methodological conclusion, a causal interpretation, a theoretical proposition or a synthesis of prior evidence. Before an experiment, “X increases Y” may be a hypothesis. After analysis, “X increased Y under these conditions” may become a result claim. The same surface sentence can therefore occupy different epistemic roles depending on when and why it is asserted.

Evidence networks should preserve this role. Turning every hypothesis into an ordinary claim node without status can make a proposed relationship look established. Research systems need explicit states such as proposed, tested, supported, unsupported, contradicted, exploratory, replicated or superseded where the domain can justify those distinctions.

3. A finding is not automatically the final claim

A finding is usually an observed or analysed result. The paper may then generalise from that result to a broader claim. A statistically significant difference in one sample is a finding. “This mechanism is generally effective in the population” is a broader inferential claim. The second requires assumptions about external validity, bias, measurement and causal interpretation that the raw result alone does not establish.

Claim-level systems should therefore connect conclusions to the intermediate findings on which they depend. Otherwise the evidence graph can create the illusion that the broad conclusion sits directly on the dataset. The missing inferential steps are where many scientific disagreements actually live.

4. Claims have scope

Scientific claims are rarely universal. They apply to populations, materials, environments, time periods, doses, measurement methods and model assumptions. “Exercise reduces blood pressure” is incomplete without enough scope to understand what population, intervention intensity, baseline condition and comparison are involved. A claim stripped of scope becomes easier to search but harder to trust.

Machine-readable claim systems must resist the temptation to compress away conditions. Atomicity should mean “small enough to reason about,” not “small enough to remove every qualifier.” The smallest truthful claim can still require several variables and constraints.

5. Qualifiers are part of the claim, not decorative uncertainty

Words such as may, likely, under these conditions, in this cohort, approximately and within confidence bounds are not rhetorical weakness. They encode the strength and boundary of the assertion. Removing them turns calibrated science into overstatement.

AI summarisation frequently compresses “was associated with” into “caused” or “suggests” into “shows”. A claim representation should preserve modality, uncertainty and study design so that the machine-readable version is no stronger than the source. Stronger wording requires stronger evidence, not smoother prose.

6. Association claims and causal claims require different evidence

“X is associated with Y” and “X causes Y” are not interchangeable relations. An observational correlation can support an association while remaining compatible with confounding, reverse causation or selection effects. A causal claim requires a design and argument that address plausible alternatives.

Evidence networks should encode the claimed relation type and the study design bearing on it. A graph that treats every statistically significant association as a causal edge will manufacture stronger science than the publications warrant.

7. Quantitative claims need units, denominators and uncertainty

A numerical claim without its measurement context is incomplete. “Risk fell by 30%” could mean relative risk reduction, absolute risk reduction, model-adjusted odds or another quantity. A value of “5” is meaningless without a unit and variable. A mean difference needs a comparison group; an effect estimate needs uncertainty or a model context.

Claim extraction from tables and figures should therefore bind number, unit, outcome, population, comparator, time and uncertainty together. Storing the headline number alone is not structured science; it is structured ambiguity.

8. Null results are claims too

A study can report that it did not detect evidence of a difference under the tested conditions. That is not the same as proving equivalence or proving that no effect exists. Null claims require careful wording because absence of statistical evidence depends on power, measurement and design.

Evidence graphs should preserve negative and null findings rather than overrepresent positive relationships. Otherwise the graph becomes publication bias in machine-readable form.

9. Methodological claims are scientifically important

Research articles make claims about methods: an assay is more sensitive, an algorithm improves accuracy, a protocol reduces variance, a model generalises better. These claims may be central to how later studies are designed even when they are not about a natural phenomenon directly.

Method claims need evidence about benchmarks, baselines, data leakage, validation sets, uncertainty and reproducibility. A claim network that only models domain findings misses the machinery that generated them.

10. Theoretical claims have evidence structures too

Not every scientific claim is a direct empirical result. Theoretical work can assert that a model explains a phenomenon, that assumptions imply a theorem, or that one conceptual framework subsumes another. Evidence may include proof, simulation, empirical correspondence, explanatory power or consistency with prior theory.

A universal evidence schema should therefore avoid assuming that every claim is a clinical trial result. Different disciplines need different evidence relations while sharing common provenance and identity principles.

11. Review claims aggregate other claims

Systematic reviews and meta-analyses produce claims from collections of studies. “The intervention probably reduces hospitalisation” is not derived from one experiment but from an evidence synthesis with inclusion criteria, effect estimates, heterogeneity, risk-of-bias judgements and possibly certainty grading.

Claim provenance for synthesis must therefore point to both the review process and the contributing studies. The claim has knowledge provenance that is inherently multi-source.

12. Guidelines and policies make action claims

A guideline may state that clinicians should use one treatment over another. This is not merely an empirical statement. It combines evidence with values, trade-offs, resource considerations and decision thresholds.

Evidence networks should distinguish descriptive claims from recommendations. Otherwise a policy conclusion can be misread as though it were a direct experimental fact.

13. A claim can be true, false, unresolved or badly specified

Verification systems often use labels such as SUPPORT, CONTRADICT and NOT ENOUGH INFORMATION. These are useful operational states, but real scientific claims can also be ambiguous, overbroad, underspecified, context dependent or based on incompatible definitions.

A system should be able to say “the claim is not precise enough to verify” rather than forcing every input into a binary truth decision. Claim quality comes before claim verification.

14. Claim atomisation is a quality-control step

A compound sentence may say that a treatment reduces symptoms, improves survival and has fewer adverse effects. One paper may support the first, contradict the second and say nothing about the third. Verification requires splitting the compound statement into independently assessable components while retaining their common context.

Good atomisation makes evidence mapping possible. Bad atomisation either leaves too many propositions bundled together or fragments meaning until the resulting claims are trivial and contextless.

Part II — Evidence is more than a citation

15. Citation is a relationship between documents; evidence is a relationship to a claim

A paper can cite another paper for historical background, a method, a definition, a criticism or a dataset. The presence of a citation therefore does not establish that the cited paper supports the surrounding claim.

An evidence relation must be evaluated at the claim level: does this source contain information that bears on the proposition? If so, does it support, challenge, qualify or merely contextualise it? This distinction is the bridge from citation graph to evidence graph.

16. Evidence has granularity

The evidence for a claim may be a full paper, one experiment, one table, one result row, one image, one paragraph or one sentence. Verification becomes more inspectable when the graph points to the smallest meaningful evidence span rather than only to the document.

Source-level citation is necessary for scholarly credit. Span-level evidence is necessary for claim-level accountability. High-quality systems keep both.

17. Evidence can be direct or indirect

Direct evidence bears on the claim through an observation or analysis closely aligned to it. Indirect evidence supports assumptions, mechanisms, measurement validity or analogous contexts. Both can matter, but they should not be treated as equivalent.

A mechanistic study in cells may support plausibility for a clinical claim without establishing patient benefit. A claim network should preserve that inferential distance.

18. Evidence is conditional on method

The same observation can carry different weight depending on how it was produced. Randomisation, blinding, calibration, sampling, preprocessing, statistical modelling and missing-data handling all influence evidential strength.

Provenance therefore includes more than “this came from Paper X.” It includes how the relevant result came to be.

19. Data are not automatically evidence

A dataset becomes evidence for a claim only through a measurement and inferential process. Raw sensor readings are not self-interpreting. A clinical database does not by itself prove a treatment effect. Evidence arises when data are connected to a question through a method.

Claim graphs should therefore connect claims not only to datasets but also to analyses and results. Skipping the method produces a provenance gap.

20. Results are not the same as evidence weight

A result can be correctly extracted and still provide weak evidence because of bias, imprecision or indirectness. “The study found X” and “X is well supported” are different claims.

Evidence networks should distinguish observed result nodes from appraisal nodes that assess reliability or applicability.

21. Replication creates independent evidence paths

If several independent studies reach compatible results using different samples or methods, the claim is supported by multiple paths. This is stronger than several papers all reanalysing the same dataset.

Graph structure can make dependence visible by connecting studies to shared datasets, laboratories, protocols or participant cohorts. Five papers are not necessarily five independent pieces of evidence.

22. Shared sources create hidden dependence

Two reviews may look independent but rely on the same small set of trials. Several machine-learning papers may benchmark on the same dataset. Multiple observational studies may use the same registry.

An evidence network should reveal shared ancestry so evidence aggregation does not count one underlying source repeatedly.

23. Evidence has publication state

A preprint, conference abstract, accepted manuscript and peer-reviewed version of record can contain related evidence at different review states. A correction may change a result. A retraction changes the trust status of the entire work or part of it.

Verification must resolve the current state before treating an older evidence span as authoritative.

24. Evidence can support part of a claim and challenge another part

A study may support the direction of an effect but not its magnitude, or support short-term benefit while challenging long-term safety. Binary document labels can hide this structure.

Claim decomposition and evidence-span alignment let systems express partial support rather than forcing a whole paper into one stance.

25. The absence of evidence is an evidence state, not a contradiction

If no retrieved source bears directly on a claim, the correct state may be unresolved. “We found no evidence” is not the same as “the claim is false.”

Automated fact-checkers should preserve this distinction. Otherwise gaps in retrieval become false negative judgements.

26. Evidence can become stale

A claim judged well supported in 2016 may be weakened by later trials, revised measurements or methodological criticism. Evidence state is time dependent.

Claim networks should support dated verification states and refresh schedules rather than storing “supported” as a permanent label.

27. Evidence networks need negative evidence and failed tests

Science advances when predictions fail as well as when they succeed. If knowledge graphs ingest only positive assertions from published papers, they reproduce publication bias and lose constraints on theory.

Negative results, failed replications and boundary conditions should be represented when they are methodologically informative.

28. Evidence networks should preserve exact source context

A sentence can look supportive when removed from the paragraph that qualifies it. A table cell can be misread without the column heading. A figure can be misunderstood without the legend.

Evidence spans should therefore preserve enough neighbouring context to interpret the result faithfully.

29. Evidence retrieval and evidence appraisal are separate stages

Retrieval asks what sources might bear on the claim. Appraisal asks how much weight those sources deserve. A system can retrieve the right paper and still make a poor verification decision if it ignores design quality.

Architectures should expose both stages so failure can be diagnosed.

30. Evidence synthesis creates a higher-order evidence object

A meta-analysis, systematic review or guideline can be treated as a derived evidence object whose provenance includes the source set, inclusion decisions, extraction rules and synthesis method.

This allows the network to represent not only first-order studies but structured summaries of evidence, without confusing the summary with the studies themselves.

Part III — Nanopublications: a compact claim-and-provenance package

31. Nanopublications make small assertions publishable

A nanopublication is a small machine-readable knowledge graph snippet treated as an independent publication. The current nanopublication community describes it as a FAIR data container for scientific results and other assertions. The goal is to make a small unit of knowledge searchable, citable, attributable and interoperable.

The idea is powerful because it separates the scientific claim from the surrounding prose while preserving who made it and how it was derived.

32. A nanopublication has an assertion graph

The assertion graph contains the main content: the proposition or structured data being published. It can be one RDF triple or a small set of triples needed to express the assertion coherently.

Atomic does not mean simplistic. If the assertion needs a measurement value, population and condition to remain truthful, the structure should preserve them.

33. A nanopublication has provenance for the assertion

The provenance graph describes how the assertion came to be. It can connect the assertion to a source publication, experiment, dataset, method, author or other generating context.

This directly answers one of the most important questions in scientific information: not merely “what is asserted?” but “what is the lineage of this assertion?”

34. A nanopublication has publication information for the nanopublication itself

The publication information graph describes the nanopublication as a published object: who created it, when it was created, how it is licensed and other metadata about the packaging act.

This distinction prevents confusion between authorship of the scientific claim and authorship of the formalisation. A person can create a nanopublication representing another researcher’s claim without becoming the originator of that claim.

35. The head graph connects the nanopublication parts

In implementation, nanopublications use a head graph that links the nanopublication to its assertion, provenance and publication-information graphs. The package becomes one addressable object.

This pattern is useful well beyond nanopublications: scientific systems need to distinguish content, provenance of content and provenance of the container.

36. Nanopublications separate scientific credit from semantic formalisation credit

A formalisation paper can represent a scientific claim from an earlier article as machine-readable RDF. The original authors deserve credit for the claim; the formaliser deserves credit for the representation.

This is a subtle but important authorship boundary. Machine-readable science creates new scholarly labour that should be attributable without stealing ownership of the underlying discovery.

37. Nanopublications can publish hypotheses, negative results and opinions

The current guidelines explicitly allow more than settled findings. Claims, hypotheses, negative results and opinions can all be represented if their type and provenance are clear.

This is healthier than a model in which every graph statement is implicitly “true.” The semantic system should preserve epistemic role.

38. Machine interpretability does not remove scientific judgement

RDF can encode a relation exactly. It cannot determine whether the relation is scientifically warranted. The choice of ontology, claim scope and evidence relation still requires domain judgement.

A syntactically valid nanopublication can represent a bad scientific claim perfectly. Formalisation improves interoperability, not truth by itself.

39. Nanopublication identity supports citation and reuse

Small claims become first-class objects when they have stable identifiers. Other nanopublications can refer to them, challenge them, derive from them or aggregate them.

This creates the possibility of scholarly discourse at claim resolution instead of only article resolution.

40. Claim-level publishing changes what can be queried

Traditional literature search asks which documents mention a concept. A claim-level system can ask which claims assert a relation, what evidence generated them, who asserted them, which claims contradict them and how those claims changed over time.

This is a transition from document retrieval to structured scientific reasoning infrastructure.

41. The basic nanopublication model is challenged by multi-source claims

Many scientific assertions are not derived from one paper. Reviews, knowledge-base entries, consensus statements and AI-generated synthesis can aggregate multiple sources.

Recent work on provenance-driven nanopublications addresses this by extending the model with knowledge provenance for assertions generated by aggregation or truth-discovery processes.

42. Knowledge provenance tracks the sources behind an aggregated claim

If an AI system concludes that Claim C is supported after analysing Papers A, B, D and E, provenance should identify the contributing sources and the process used to combine them.

Without this lineage, the aggregated claim looks like a direct assertion with no visible reasoning history.

43. Trust networks emerge when claims share provenance

Several claims may derive from the same evidence source, laboratory, database or extraction model. Provenance links reveal these dependencies.

A claim network can therefore reveal not only agreement but common ancestry. Apparent consensus may be weaker when every supporting claim ultimately traces to one experiment.

44. Versioning is essential for claim-level publication

A claim may be corrected, narrowed or replaced. The old claim should not simply vanish if it was previously cited or used in a model.

Version relationships let later systems understand which assertion state was active at a particular time and which newer state superseded it.

45. Retraction at claim level is different from deleting a claim

If a claim was published and later withdrawn, historical systems may need to know that it existed and influenced downstream work. Deletion destroys that history.

A retracted or invalidated state preserves provenance while warning current users.

Part IV — Provenance: how scientific knowledge came to be

46. Provenance is the history of production

Provenance answers questions such as: Who produced this result? Which dataset was used? Which analysis generated it? Which source was it derived from? Which software version participated? When did the activity occur?

In scientific reasoning, provenance is not bureaucratic metadata. It is part of the evidence required to judge reliability and reproducibility.

47. W3C PROV separates entities, activities and agents

The W3C PROV data model and PROV-O ontology provide a general framework for interoperable provenance. An entity can be a dataset or result; an activity can be an analysis or experiment; an agent can be a person, organisation or software agent responsible for actions.

This separation allows a graph to represent how an entity was generated by an activity and attributed to an agent rather than storing a flat source string.

48. wasDerivedFrom is not the same as wasGeneratedBy

A result may be generated by an analysis activity and derived from an input dataset. These are different provenance relations.

Fine-grained provenance lets a user reconstruct the causal production chain instead of receiving an undifferentiated “source” field.

49. Attribution provenance and derivation provenance answer different questions

“Who should receive credit or responsibility?” and “From what information was this derived?” are not the same question. A data curator may publish a cleaned dataset derived from raw measurements produced by another team.

Evidence networks should preserve both attribution and derivation.

50. Provenance can be prospective or retrospective

Prospective provenance describes the planned workflow: the protocol, intended analysis or pipeline. Retrospective provenance records what actually happened: which files, parameters and versions were used.

Reproducibility benefits from both. The plan explains intention; the execution record explains the realised computation.

51. Provenance granularity should match the decision

Not every claim needs a trace of every CPU instruction. A high-level literature claim may need publication and evidence-span provenance. A computational claim may require dataset version, code commit, model checkpoint and random seed.

Provenance should be detailed enough to audit the inference without overwhelming the system with irrelevant events.

52. Provenance should be immutable enough to audit and flexible enough to correct

A provenance record should preserve what was known or asserted at the time while allowing later correction of mistakes. The usual solution is additive history: preserve the earlier record, attach a correction event and expose the current preferred view.

Silent overwriting makes scientific history impossible to reconstruct.

53. Provenance can expose circular evidence

Two reviews may cite each other while both ultimately rely on the same primary experiment. An AI answer may cite a secondary summary that cites another summary that returns to one original study.

A provenance graph lets the verifier follow the chain to primary evidence and detect when apparent multiplicity is actually citation recycling.

54. Provenance can reveal model-generated knowledge

If an LLM extracted a claim from a paper, the graph should identify the model, version, prompt or extraction configuration where practical, plus the source passage used.

Model-generated structure should not be indistinguishable from author-deposited metadata or expert curation.

55. Provenance is what makes automated correction feasible

If one extraction model is found to have a systematic error, provenance allows the system to identify which graph edges were produced by that model and recompute them.

Without production lineage, correcting a model becomes a manual hunt through an opaque graph.

Part V — Support, contradiction and typed scientific discourse

56. Support is not the same as citation

CiTO, the Citation Typing Ontology, exists because citations have functions. A citation can extend prior work, use a method, provide data, support, dispute or discuss. The current CiTO version, updated in September 2026, explicitly supports factual and rhetorical characterisation of citation intent.

Evidence systems should use typed relations or equivalent semantics rather than infer that every outgoing reference is positive support.

57. Support can mean entailment, consistency or corroboration

A paper may directly entail a claim through a result, be merely consistent with the claim, or provide independent corroboration. These are not identical forms of support.

Claim verification systems can simplify them operationally, but high-trust evidence graphs should preserve the distinction where it changes interpretation.

58. Contradiction can mean logical conflict or empirical tension

Two claims logically contradict when they cannot both be true under the same conditions. Scientific studies often disagree more subtly: different populations, doses, measurements or model specifications produce different results.

A contradiction edge should therefore ask whether the scope conditions align. Otherwise normal heterogeneity becomes false conflict.

59. Challenge is a useful relation broader than contradiction

A paper may challenge a claim by identifying bias, questioning a mechanism or showing that the result fails under a boundary condition without logically contradicting every version of the proposition.

Evidence networks benefit from relations such as supports, challenges, refines, limits, replicates and fails-to-replicate rather than only support/contradict.

60. Replication claims should connect to the target claim and target method

A replication attempts to test whether an earlier result can be reproduced under aligned or deliberately varied conditions. The graph should connect the replication not only to the earlier paper but to the specific claim and methodological elements being tested.

This makes “failed replication” interpretable: failed under which conditions, with which differences?

61. Citation context contains evidence about citation intent

The surrounding text of an in-text citation often reveals whether the author is using, extending, disputing or merely mentioning the cited work.

Claim networks can extract citation contexts as evidence for typed citation relations while preserving the original sentence so the classification can be checked.

62. A claim can have a support ratio without having a truth score

Counting supporting and challenging sources can summarise the network, but the ratio should not be interpreted as probability of truth unless a defensible statistical model justifies that meaning.

Ten low-quality dependent studies do not automatically outweigh two strong independent studies. Evidence aggregation needs weighting and dependence structure.

63. Evidence weight should be task-specific

A clinical guideline may prioritise randomised trials and systematic reviews. A rare-safety signal may depend on observational data. A mathematical claim may require proof. A historical claim may rely on primary archival documents.

One universal evidence hierarchy would distort disciplines. Claim networks need domain-aware appraisal models.

64. Independence matters more than raw source count

Evidence from five laboratories using independent samples may be more informative than five analyses of the same dataset. Shared methods and assumptions create correlated error.

Provenance graphs can expose common dependencies so aggregation models do not assume independence incorrectly.

65. Source quality and evidence relevance are separate axes

A rigorous study can be irrelevant to the specific claim. A directly relevant study can be methodologically weak.

Verification should evaluate both: Does the source bear on the proposition, and how much trust should its result receive?

66. Evidence graphs can make disagreement productive

Instead of flattening conflicting papers into “mixed evidence,” a graph can expose which variables separate them: population, time, dose, instrument, preprocessing, outcome definition or statistical model.

The disagreement becomes a map of boundary conditions that can generate the next experiment.

67. Multi-source assertions need explicit aggregation provenance

Recent provenance-driven nanopublication work extends ordinary assertion provenance to capture claims produced by aggregating several knowledge sources. This matters for truth-discovery systems, knowledge bases and AI synthesis.

The output claim should identify which sources contributed and which algorithm or rule combined them.

68. Consensus is a property of a defined evidence set

“The literature agrees” is incomplete without saying which literature was searched, which studies were included and how disagreement was handled.

A claim network can make consensus reproducible by attaching it to a dated source set and synthesis process.

69. Minority evidence should remain visible

Low-frequency contradictory findings can reveal subgroups, measurement problems or emerging paradigm shifts. Removing them because they disagree with the majority weakens scientific self-correction.

Graphs should allow ranking by weight while preserving dissenting paths.

70. Retracted evidence should remain in the discourse graph with altered status

A retracted paper may have generated support relations in earlier literature. Deleting the node hides how the field evolved.

The graph should preserve historical support edges while marking the source as retracted and preventing current verification from treating it as ordinary clean evidence.

Part VI — Scientific claim verification as a pipeline

71. Verification begins by normalising the claim without strengthening it

The system should rewrite the claim into a clear proposition while preserving modality, scope and quantities. It must not silently turn “may be associated” into “causes.”

The normalised claim becomes the retrieval target and should remain traceable to the original wording.

72. Claim decomposition identifies independently verifiable components

Compound claims should be split when different evidence can bear on each component. Entity, relation, quantity, population, condition and time can be extracted as structured fields.

The decomposition becomes a checklist: every component should receive evidential support or an explicit unknown state.

73. Retrieval should search for evidence, not merely similar language

Semantic similarity is useful for candidate discovery but can return papers discussing the topic without testing the claim. Evidence retrieval should combine lexical, semantic, citation and graph routes.

A strong verifier asks which studies can actually change the claim’s status.

74. Source universe must be explicit

Did the verifier search PubMed, Crossref, Semantic Scholar, OpenAlex, a curated corpus or the open web? Coverage differences determine what evidence can be found.

“Verified against the literature” is too vague. High-standard systems report the retrieval universe and date.

75. Evidence selection should identify exact rationales

SciFact operationalised scientific claim verification by pairing claims with evidence rationales from research abstracts and labels such as SUPPORT or CONTRADICT. The key idea is that the verifier should point to the sentences that justify the stance.

This makes evaluation possible: a model can retrieve the right document but choose the wrong rationale, or choose the right rationale but infer the wrong label.

76. Evidence stance should be evaluated relative to the exact claim

The same paper can support one claim and contradict another. The label belongs to the claim–evidence relation, not to the document globally.

This is why evidence networks model edges, not merely paper categories.

77. Verification requires context alignment

A result in mice may not verify a claim about humans. A result at a high dose may not verify a low-dose claim. A retrospective association may not verify a causal statement.

Context alignment compares population, intervention or exposure, outcome, measurement, time and study design before assigning stance.

78. Verification should check publication state

Before final judgement, resolve whether the source is a preprint, corrected article, retracted article or superseded version.

Otherwise the system may accurately extract evidence from a source that is no longer scientifically current.

79. Verification should distinguish primary evidence from commentary

A review can summarise evidence accurately, but high-stakes verification may need to return to primary studies. Secondary sources can also misstate or selectively cite.

The graph should show whether support is direct from primary data or mediated through another synthesis.

80. Verification should retrieve disconfirming evidence deliberately

A search optimised only for semantic similarity to the claim can preferentially retrieve confirmatory language. A robust verifier should also search for negations, failed replications, criticism, boundary conditions and alternative explanations.

Scientific verification is adversarial retrieval against one’s own preferred answer.

81. Verification needs a stopping rule

The scientific literature can be enormous. A verifier cannot read everything for every query. It needs a principled stopping condition based on retrieval saturation, source quality, time budget or task risk.

High-stakes claims require a broader search than exploratory questions. The system should expose that trade-off.

82. Verification outputs should separate observed evidence from judgement

OBSERVED:
  Source A reports result R under conditions C.

RELATION:
  R supports component 1 of Claim X.

APPRAISAL:
  Source A has these strengths and limitations.

SYNTHESIS:
  Sources A, B and D support; C challenges.

JUDGEMENT:
  Claim X is currently supported with stated uncertainty.

Separating these layers prevents a model from hiding interpretation inside what looks like a sourced fact.

83. Verification should produce a claim receipt

CLAIM RECEIPT

Original claim:
Normalised claim:
Scope:
Date verified:
Source universe:
Search strategy:
Evidence sources:
Supporting spans:
Challenging spans:
Publication states:
Method-quality notes:
Known dependence:
Synthesis rule:
Current judgement:
Uncertainty:
Next review date:

The receipt turns verification into an auditable research object.

84. SciTrue illustrates the move toward source-level accountability

SciTrue, presented at EACL 2026, was designed to address unverifiable attributions, shallow evidence mapping and hallucinated citations in AI-assisted scientific claim verification. Its core design principle is explicit mapping from claim components to verifiable scientific sources.

The important lesson is architectural: a verifier should not merely output a verdict with citations. It should show how each component of the claim is grounded.

85. General-purpose web fact-checking and scientific verification differ

Scientific claims often require interpreting study design, methods, uncertainty and evolving publication states. A reputable news article may be useful context but is not the same evidential object as a trial or dataset.

Scientific verification therefore needs domain-aware source selection and appraisal.

86. Verification can be correct for the wrong reasons

A model may predict the correct SUPPORT label while citing an irrelevant sentence. Benchmark evaluation should therefore score evidence retrieval and stance separately.

Answer accuracy without evidence fidelity is fragile.

87. Verification can be wrong even with perfect retrieval

The right sources can be retrieved and the model can still misunderstand statistical significance, population scope, negation or causal language.

Reasoning evaluation should therefore be separated from retrieval evaluation.

88. “Not enough information” is a success state when evidence is insufficient

Scientific systems must be rewarded for calibrated abstention. Forcing an answer encourages hallucinated certainty.

A verifier should explain whether uncertainty comes from missing evidence, conflicting evidence, claim ambiguity or low-quality evidence.

89. Claim verification is dated

A verdict should include when the evidence search was performed. New studies, corrections and retractions can change the state.

“Supported as of 18 September 2026 under this evidence set” is scientifically better than “true.”

90. Verification should be re-runnable

Saved queries, source identifiers, graph versions and evidence spans allow later systems to refresh the judgement instead of starting from zero.

Re-verification becomes a maintenance workflow.

Part VII — Contradiction, uncertainty and scientific disagreement

91. Many apparent contradictions are scope mismatches

One study may involve children and another older adults; one high dose and another low dose; one laboratory outcome and another clinical outcome. Results can differ without logical contradiction.

Before labelling evidence conflicting, align the conditions under which each claim is asserted.

92. Measurement differences can create disagreement

Two studies can measure “engagement,” “depression” or “performance” using different instruments. The variable names match while the operational definitions differ.

Evidence graphs should connect claims to measurement definitions when those definitions affect comparability.

93. Statistical model choices can create apparently different findings

Adjusted and unadjusted models, different covariates, transformations and missing-data procedures can change estimates.

Conflict analysis should inspect analysis provenance instead of assuming one study must be wrong.

94. Temporal context can resolve conflict

A relationship can change after a policy, variant, technology or treatment practice changes. Evidence from different periods may both be correct for their own context.

Temporal qualifiers belong inside the claim model.

95. A failed replication is evidence about reproducibility, not an automatic refutation

The replication may differ in power, protocol or population. Failure to reproduce should trigger analysis of conditions and uncertainty.

Graph relations can encode failed-to-replicate without forcing a global “false” state.

96. Contradictions can expose hidden moderators

When support and challenge cluster by population, method or environment, the network suggests a moderator that earlier broad claims omitted.

Disagreement becomes a discovery engine for better claim scope.

97. Uncertainty should be decomposed

• Sampling uncertainty.
• Measurement uncertainty.
• Model uncertainty.
• Bias uncertainty.
• Evidence-coverage uncertainty.
• Publication-state uncertainty.
• Claim-parsing uncertainty.
• Entity-resolution uncertainty.

One undifferentiated confidence number hides where the system is fragile.

98. Scientific disagreement should preserve source independence

If five challenging papers all originate from one laboratory and two supporting papers come from independent multi-centre studies, raw paper count misleads.

Provenance enables dependence-aware interpretation.

99. Disagreement graphs should retain minority paths

A later paradigm shift often begins as minority evidence. Pruning low-frequency conflict because it reduces interface complexity can erase early warning signals.

Rank by weight, but preserve inspectability.

100. The strongest output is often a conditional claim

Instead of “X works” or “X does not work,” evidence may support “X improves outcome Y in population P under condition C, while evidence is weaker outside that setting.”

Scientific reasoning improves when the system learns to narrow the claim rather than merely choose a side.

Part VIII — Machine extraction of claims and evidence

101. Claim extraction is not summarisation

Summarisation produces concise text. Claim extraction identifies propositions that can be individually assessed. A summary can merge several claims or omit qualifiers for readability.

Structured extraction needs a schema and fidelity constraints.

102. Named entity recognition identifies the nouns of the scientific graph

Entities may include diseases, genes, materials, interventions, populations, instruments, datasets and methods. Correct entity linking allows claims to connect across papers.

Entity recognition without disambiguation can attach the right name to the wrong concept.

103. Relation extraction identifies the verb of the claim

Relations such as increases, inhibits, associated-with, predicts, uses-method and contradicts create structured connections.

Relation labels must reflect the source wording. An association sentence should not become a causal predicate.

104. N-ary claims require more than subject–predicate–object

“Drug X reduced outcome Y by 18% versus placebo in adults over 12 weeks” contains comparator, population, magnitude and time. Compressing it to Drug X reduces Y discards scientifically essential dimensions.

Claim schemas need slots, qualifiers or event nodes for multi-argument assertions.

105. Rosetta-style statement patterns reduce modelling friction

Recent work on Rosetta Statements aims to make FAIR knowledge graph construction easier through user-centred statement templates. Each statement version can be represented as a nanopublication.

The broader lesson is that scientists should not need to become ontology engineers to publish structured claims. Templates can preserve semantic discipline while hiding unnecessary implementation complexity.

106. Table extraction is essential because many results are not in prose

Key effect sizes, sample counts and subgroup results often live in tables. A text-only claim extractor can miss the strongest evidence.

Table extraction must preserve row and column headers, units, footnotes and statistical annotations.

107. Figure extraction is riskier because visual context matters

Plots may encode trends, uncertainty bars and subgroup differences that are not fully described in captions. Extracting a claim from a figure requires visual interpretation and should preserve the image region or underlying data where available.

Visual inference deserves its own confidence and provenance.

108. LLMs can improve extraction but also invent structure

Language models can understand complex scientific sentences, resolve references and map text into schemas. They can also hallucinate entities, strengthen modality or fill missing slots with plausible values.

Schema-constrained output, evidence-span citation and validation rules are therefore non-negotiable in high-trust extraction.

109. Extraction should retain the raw source passage

Every model-generated claim should point to the exact text, table or figure from which it was extracted.

The source span is the repair interface when the model is wrong.

110. Claim normalisation should not erase author uncertainty

If the paper says “our results suggest,” the structured claim should not become unqualified. Modality should be represented as a field or relation property.

Machine-readable certainty must never exceed human-authored certainty without new evidence.

111. Extraction models need domain-specific evaluation

A model that performs well on biomedical claims may fail on materials science, economics or education because entities, relation types and evidence structures differ.

Evaluation should use representative domain corpora rather than assuming one benchmark generalises universally.

112. Inter-annotator disagreement is useful information

If experts disagree on whether a sentence contains one claim or two, or whether an evidence span supports a proposition, the task may be intrinsically ambiguous.

Gold standards should preserve difficult cases rather than force artificial unanimity.

113. Automated extraction needs drift monitoring

Scientific language changes. New model names, biomarkers, measurement practices and reporting standards appear.

Extraction precision can decline silently over time, so production systems need recurring audits and update sets.

114. Claim extraction should be reversible

If a model version is later found defective, the graph should identify its outputs and regenerate them from source passages.

Transformation provenance makes this possible.

Part IX — AI scientific reasoning over evidence networks

115. Retrieval-augmented generation is not verification by itself

RAG gives a language model retrieved context. It does not guarantee that the context is relevant, that the model interprets it correctly, or that conflicting evidence was retrieved.

Verification requires explicit claim–evidence mapping and stance assessment beyond ordinary RAG.

116. Claim-first retrieval is more disciplined than topic-first retrieval

Instead of asking “find papers about X,” a claim-aware system extracts the proposition and searches for evidence that could support or challenge each component.

This narrows retrieval from topical relevance to evidential relevance.

117. Evidence graphs can act as an operational state for research agents

Recent autonomous-research work such as EviGraph represents problems, gaps, hypotheses, experiments, findings and claims in a typed evidence graph. The graph is maintained during the research process rather than produced only after the manuscript is written.

This architecture is important because it lets the system check whether claims remain connected to validated evidence before drafting conclusions.

118. AI reasoning should traverse evidence chains, not merely similarity clusters

Semantic search can find similar papers. An evidence chain shows how a claim is linked to experiments, results and sources.

Scientific reasoning benefits when the model receives structured paths plus the original source passages.

119. A graph path explains retrieval, not necessarily scientific inference

Claim → supported by → Study A → uses → Dataset D is a provenance route. It does not prove that Study A’s design is sufficient or that Dataset D is unbiased.

AI systems should distinguish graph connectivity from evidence appraisal.

120. AI should verify every generated citation

A hallucinated DOI or fabricated paper can make an otherwise plausible answer unusable. Generated references should be resolved against authoritative metadata before publication or high-stakes use.

Citation verification belongs before answer release, not after a user complains.

121. AI should verify every claim component, not just attach one source to a paragraph

A paragraph may contain five factual propositions. One citation at the end can support only two. Claim-level attribution exposes this mismatch.

SciTrue’s source-level accountability direction reflects this need for component-level grounding.

122. AI should retrieve challenge evidence intentionally

Evidence networks let a system perform a counter-search: after building a support case, look for challenge, contradiction, retraction, failed replication and boundary evidence.

This is an automated form of scientific adversarial checking.

123. AI should know when the graph is incomplete

Missing edges can reflect missing metadata rather than absence of a real relation. A dataset may be used without being formally linked; a paper may be relevant but unindexed.

The system should say “not represented in this graph” instead of “does not exist.”

124. AI should not collapse provenance into one citation

An aggregated claim may depend on several studies plus a synthesis rule. Citing one review may conceal the source set.

For consequential claims, the answer should expose the evidence bundle or at least provide a route to inspect it.

125. AI should separate claim extraction confidence from claim truth confidence

The model can be highly confident that a paper says “X was associated with Y” while the scientific evidence for the association remains weak.

Extraction confidence concerns interpretation of the source; evidential confidence concerns the world claim. They are different axes.

126. AI should preserve uncertainty through generation

If evidence is mixed, the final prose should remain mixed. A summariser should not convert three supportive and two challenging studies into “research proves.”

Generation should be constrained by the evidence state.

127. AI can use claim receipts as compact reasoning interfaces

Instead of passing dozens of full papers to a model, a system can assemble structured claim receipts with evidence spans, provenance and quality notes, then provide links to full sources for verification.

This reduces context waste while keeping reasoning inspectable.

128. AI agents should maintain evidence state throughout a workflow

If an autonomous research agent changes a hypothesis after an experiment, downstream claims and manuscript sections should be invalidated or regenerated. The evidence graph can encode these dependencies.

Research state becomes a directed system rather than a sequence of chat messages.

129. AI should not promote hypotheses to findings automatically

A research agent may generate a plausible hypothesis and then write about it fluently. Until an experiment or evidence source tests it, the graph should keep it in a hypothesis state.

State transitions should require explicit evidence events.

130. AI should not promote findings to general claims automatically

One experiment can yield a finding. Generalising to a broader scientific claim requires an inferential step that may need replication or synthesis.

Evidence networks can force the system to record that step instead of hiding it in prose.

131. AI scientific reasoning should be correction-aware

If a key source is retracted after an answer is generated, the claim state should be recomputed when the system next refreshes.

Long-lived AI knowledge bases need update propagation, not one-time indexing.

132. The safest AI answer contains both conclusion and audit path

CONCLUSION
  → exact claim
  → supporting evidence
  → challenging evidence
  → evidence quality
  → publication states
  → provenance
  → unresolved gaps
  → dated verification
  → source-return links

The answer becomes a navigable evidence product rather than a polished endpoint.

Part X — Publishing claim–evidence structure

133. Scientific papers already contain an implicit claim graph

The introduction motivates a problem, methods produce observations, results report findings and discussion sections generalise them. Citations support background and argument.

Claim-aware publishing makes these relationships explicit rather than leaving every reader or AI system to reconstruct them independently.

134. Authors could publish a claim map alongside the article

A claim map can list major claims, source result sections, datasets, analysis code and external evidence used for contextual claims.

This would not replace the article; it would provide a machine-readable table of contents for scientific reasoning.

135. Peer review can evaluate claim–evidence alignment explicitly

Reviewers often already ask whether conclusions overreach the data. A claim graph turns that judgement into a visible structure: which claim depends on which result, and where is the inferential jump?

Peer review can annotate weak edges rather than only comment on paragraphs.

136. Corrections can target affected claims

If an error changes one figure but not the rest of the paper, a claim network can identify which downstream conclusions depend on that figure.

Correction becomes dependency-aware rather than document-wide guesswork.

137. Retractions can propagate through affected evidence networks

If a retracted paper was a major support source for several review claims, systems can flag those claims for re-evaluation.

This is one of the strongest practical reasons to build structured evidence relations.

138. Crossref’s Research Nexus provides the broader relationship infrastructure

Crossref’s Research Nexus vision describes a reusable open network connecting research organisations, people, things and actions. Its documentation emphasises relationships for research integrity, reproducibility, assessment and discoverability.

Claim networks sit naturally inside this direction: the richer the relationships among outputs, data, software, protocols and corrections, the easier it becomes to build evidence lineage.

139. Publishing structured claims should remain optional and incremental

Requiring every researcher to encode every sentence as RDF would create resistance and low-quality metadata. The useful starting point is major claims and high-value relations.

Structured publishing should reduce repeated extraction work, not add ceremonial bureaucracy.

140. Templates can move semantic structure upstream

If authors select statement templates or structured result forms during writing, later systems do less uncertain extraction.

Upstream semantics are usually more reliable than downstream inference because the researcher still knows what the claim means.

Part XI — Systematic reviews, meta-analysis and evidence synthesis

141. A systematic review is already an evidence network in prose

Reviewers collect studies, link reports to underlying studies, assess bias, extract outcomes and combine compatible results. The final conclusion depends on a complex evidence structure.

Representing that structure explicitly allows readers to inspect how each synthesis claim was constructed.

142. Study identity should sit below report identity

One study can generate a protocol, conference abstract, main paper and follow-up analysis. Treating each report as independent evidence inflates support.

Evidence graphs should link multiple reports to one study entity where appropriate.

143. Outcome harmonisation is claim normalisation

Different studies may measure similar concepts with different instruments. Meta-analysis requires deciding whether outcomes are sufficiently compatible.

A claim network can preserve the original measures while linking them to a harmonised analytical construct.

144. Risk-of-bias judgements should be nodes or annotations, not invisible reviewer memory

Evidence weight depends on study limitations. Structured bias assessments can attach to studies and flow into synthesis claims.

The reader can then see why two studies with similar results received different weight.

145. Meta-analysis creates a derived result with mathematical provenance

The pooled effect depends on included studies, extracted values, transformations, weighting model and heterogeneity assumptions.

Those inputs should be preserved so the synthesis result can be recalculated when a source study changes.

146. Living reviews require continuously maintained claim states

When new trials arrive, evidence networks can identify which synthesis claims depend on the updated evidence set and need recomputation.

This is more scalable than treating every review update as a completely new document.

Part XII — Research integrity through claim–evidence auditing

147. Unsupported conclusion inflation becomes visible

If a paper’s abstract claims broad effectiveness but the graph shows only a small exploratory result, the weak edge is visible.

Claim–evidence auditing can help editors and reviewers detect overstatement.

148. Citation laundering becomes traceable

A secondary source may assert a claim and cite another review, which cites an original study that did not actually support the proposition. Repeated citation can create false authority.

Evidence provenance lets the verifier follow the chain to the primary source.

149. Cherry-picking appears as asymmetric evidence selection

If a review claim links only to supportive studies while equally relevant challenging studies exist in the searched corpus, the evidence network exposes the omission.

Integrity analysis can compare the available evidence neighbourhood with the cited evidence set.

150. P-hacking cannot be diagnosed from a claim graph alone

Evidence networks can show multiple outcomes, analyses and versions, but statistical misconduct requires deeper methodological evidence.

The graph should route investigation, not pronounce misconduct based on pattern alone.

151. Image and data concerns can be attached to affected claims

If an image is duplicated or a dataset is corrupted, downstream claims relying on that evidence can be flagged.

This creates a blast-radius model for integrity corrections.

152. Retraction notices should specify affected claims when possible

Some retractions invalidate the entire article; others stem from a specific result. More granular correction metadata would allow evidence networks to update precisely.

Current publishing systems are still mostly document-level, but claim-level infrastructure makes finer repair conceivable.

153. Integrity signals are not verdicts

Unusual citation clusters, missing provenance or unsupported claims can indicate areas for review. They do not prove misconduct.

Systems should preserve due process and clearly distinguish automated anomaly detection from human findings.

154. Auditability should be designed before controversy

The best time to create provenance is when research is produced, not years later when a dispute arises.

Structured claim and evidence records make later investigation more reliable because they reduce dependence on reconstructed memory.

Part XIII — Standards and interoperability

155. RDF provides a shared graph representation

Resource Description Framework represents statements as triples using resolvable identifiers. Nanopublications use RDF because it supports linked data and named graphs.

RDF is a representation technology, not an epistemic guarantee. The ontology and provenance still determine scientific meaning.

156. PROV-O provides interoperable provenance vocabulary

W3C PROV-O is a stable Recommendation for expressing provenance with entities, activities, agents and their relationships.

Domain evidence ontologies can extend PROV rather than reinvent provenance from scratch.

157. CiTO provides typed citation intent

CiTO’s current 2026 version formalises factual and rhetorical citation characterisations such as supports, disputes, extends and uses-method relations.

This makes citation graphs more semantically useful for evidence analysis.

158. EVI extends provenance toward evidence correctness

The Evidence Graph Ontology released in 2026 extends PROV concepts to describe evidence supporting the correctness of findings in biomedical publications.

This reflects a broader trend: provenance systems are becoming evidence systems when they explicitly model how findings are justified.

159. Persistent identifiers connect claims to durable scholarly entities

DOI, ORCID, ROR and related identifiers reduce ambiguity when claims cite works, people or organisations.

Claim-level systems should reuse existing PID infrastructure rather than invent local strings for globally identified entities.

160. Crossref relationships connect broader research objects

Crossref’s Research Nexus vision emphasises versions, corrections, translations, data, supplements, grants, reviews and other relationships among research objects.

Claim–evidence networks become more powerful when they can traverse this existing relationship infrastructure.

161. Interoperability requires relation definitions, not matching labels alone

Two ontologies can both use “supports” while meaning different things. Mapping needs semantic alignment: domain, range, conditions and evidence expectations.

Vocabulary reuse is valuable only when meaning is preserved.

162. Claim identifiers should survive platform migration

A claim network can become long-lived scholarly infrastructure only if identifiers remain stable when databases change.

Persistent claim identity enables citation, correction and version history.

Part XIV — Evaluation gym: how to test a claim–evidence system

163. The compound-claim test

Give the system a sentence containing three propositions with different evidence states. Verify that it decomposes them without losing shared scope.

This tests atomisation.

164. The association-versus-causation test

Use an observational study reporting association and ask the system to verify a causal claim. It should reject the strengthening or lower confidence.

This tests relation fidelity.

165. The qualifier-preservation test

Use a source saying “may modestly improve” and verify that the extracted claim does not become “improves.”

This tests modality fidelity.

166. The contradictory-evidence test

Provide a literature set with strong support and strong challenge. The system should retrieve both and explain the scope of disagreement.

This tests confirmation-bias resistance.

167. The shared-data dependence test

Provide several papers derived from one dataset and another independent replication. The system should not count every paper as independent support.

This tests provenance-aware aggregation.

168. The retraction-propagation test

Retract a source node and verify that dependent claim states are flagged for refresh.

This tests correction awareness.

169. The preprint-version test

Give the system preprint and final versions with changed conclusions. It should preserve version identity and prefer current state.

This tests version reasoning.

170. The evidence-span test

Require the model to identify the exact sentences or table cells that support the claim. Score whether the stance can be justified from those spans alone.

This tests source-level accountability.

171. The source-universe test

Run the same claim against different databases and compare verdicts. Large changes reveal coverage sensitivity.

This tests retrieval boundary awareness.

172. The stale-verdict test

Add a major new study after initial verification. The system should detect that the evidence set changed and refresh the judgement.

This tests maintenance architecture.

173. The missing-evidence test

Ask about a claim for which the corpus contains no relevant evidence. The correct output is unresolved, not fabricated support or contradiction.

This tests abstention.

174. The adversarial-citation test

Provide a cited paper that does not actually support the surrounding sentence. A strong system follows the citation and detects the mismatch.

This tests citation laundering resistance.

175. The table-only-evidence test

Place the decisive result in a table while the prose remains vague. The verifier should retrieve or inspect structured table evidence.

This tests multimodal evidence coverage.

176. The domain-transfer test

Evaluate claim extraction and stance classification in a field different from the training data.

This reveals whether the schema and model generalise or merely memorise domain conventions.

Part XV — Practical architecture for a trustworthy evidence network

177. Layer 1: claim intake

Preserve the original claim text, author, context, source and date. Do not begin by rewriting it.

The original is the reference point for every later transformation.

178. Layer 2: claim parsing and normalisation

Identify propositions, entities, relation type, quantities, conditions, modality and time. Produce independently verifiable claim components.

Store the transformation provenance and model version.

179. Layer 3: authoritative entity resolution

Resolve papers to DOI or other PIDs, researchers to ORCID when available, organisations to ROR and domain entities to appropriate ontologies.

Ambiguous resolution should remain ambiguous until enough evidence exists.

180. Layer 4: evidence retrieval

Search lexical, semantic, citation and graph channels. Record source universe, date and strategy.

Retrieve challenge evidence deliberately.

181. Layer 5: evidence-span extraction

Identify exact sentences, tables, figures or structured results bearing on the claim.

Keep enough surrounding context to preserve meaning.

182. Layer 6: stance classification

Classify each evidence relation as support, challenge, contradiction, neutral context or insufficient. Preserve partial support where possible.

Store confidence separately from evidence quality.

183. Layer 7: method and source appraisal

Assess study design, bias, directness, precision, independence and publication state using domain-appropriate rules.

Appraisal modifies weight, not the literal content of the result.

184. Layer 8: provenance graph

Connect claims, evidence spans, datasets, methods, analyses, publications, agents and transformation activities.

Use standards such as PROV-O where appropriate.

185. Layer 9: evidence aggregation

Combine evidence with explicit rules that account for dependence and quality. Do not convert source count directly into truth probability.

Preserve minority and conflicting paths.

186. Layer 10: dated claim state

Produce a current judgement such as supported, challenged, unresolved or context-dependent, together with uncertainty and date.

The state is a snapshot, not an eternal property.

187. Layer 11: publication-state monitor

Track corrections, retractions, new versions and new evidence. Trigger re-verification when high-impact dependencies change.

Maintenance is part of verification.

188. Layer 12: human challenge and correction

Allow qualified users to report misparsed claims, wrong evidence spans, missed studies and incorrect entity links.

Corrections should be versioned and attributed.

Part XVI — Failure library

Failure	What breaks	Repair
Sentence = claim	Compound propositions remain tangled	Semantic claim decomposition
Citation = support	Background references become evidence	Typed claim–evidence relations
Similarity = evidence	Topical papers are mistaken for verifying sources	Evidence-span alignment
Association = causation	Claims are strengthened beyond design	Preserve relation semantics
Qualifier dropped	Uncertain science becomes certainty	Model modality explicitly
One citation covers paragraph	Unsupported subclaims hide inside prose	Claim-level attribution
Document stance	One paper is treated as globally supportive or contradictory	Edge stance per claim
Paper count = evidence weight	Dependent studies are overcounted	Provenance-aware aggregation
Retraction deletes node	Historical influence disappears	Preserve node, change status
Model confidence = truth confidence	Extraction certainty becomes scientific certainty	Separate confidence layers
No exact evidence span	Verification cannot be challenged	Attach sentence/table/figure evidence
Only supportive retrieval	Confirmation bias is automated	Counter-search for challenge evidence
No source universe	Coverage limits disappear	Record databases and search date
Multi-source claim lacks lineage	Aggregated conclusion appears source-free	Knowledge provenance
Old verdict stored permanently	Science cannot self-correct	Dated states and refresh triggers
LLM extraction overwrites source	Model interpretation becomes canonical text	Preserve raw source and transformation provenance

189. Failure analysis should happen before deployment

Teams should build an adversarial library of ambiguous claims, mixed evidence, retractions, version conflicts and citation mismatches before the system reaches users.

Scientific claim verification fails in plausible edge cases more often than in obvious nonsense.

190. The most dangerous failure is fluent unsupported synthesis

An AI answer can sound balanced, cite real papers and still misrepresent what those papers show. Surface professionalism hides evidence misalignment.

Source-level accountability is the antidote.

Part XVII — Practical checklists

191. Researcher checklist

• State major claims with population, condition and uncertainty.
• Separate findings from broader interpretation.
• Link major claims to exact results.
• Deposit data and code with persistent identifiers where appropriate.
• Use ORCID and organisation identifiers in workflows.
• Preserve negative and null findings.
• Correct claim wording when evidence changes.
• Avoid citing reviews as substitutes for primary evidence when primary evidence is the question.
• Distinguish association from causation.
• Keep claim scope no broader than the evidence.

192. Publisher checklist

• Capture persistent identifiers and reference relationships.
• Support data, software and protocol links.
• Expose correction and retraction status machine-readably.
• Consider structured major-claim metadata.
• Preserve version relationships.
• Use provenance for enriched or extracted metadata.
• Make claim corrections additive and traceable.
• Keep article-level and claim-level authorship roles distinct.

193. Systematic reviewer checklist

• Identify study families across multiple reports.
• Attach synthesis claims to included studies.
• Preserve risk-of-bias judgements.
• Record outcome harmonisation decisions.
• Expose conflicting studies rather than hiding them in pooled estimates.
• Version the evidence set.
• Refresh claim states when new evidence arrives.

194. AI claim-verifier checklist

• Preserve original claim text.
• Decompose compound claims.
• Retain qualifiers and modality.
• Resolve entities and publication versions.
• Search a declared evidence universe.
• Retrieve exact evidence spans.
• Search for support and challenge.
• Assess context alignment.
• Separate retrieval, stance and appraisal.
• Check retractions and corrections.
• Produce a dated claim receipt.
• Abstain when evidence is insufficient.

195. Evidence-graph builder checklist

• Define claim schema by domain.
• Reuse PROV-O for provenance where appropriate.
• Reuse citation semantics such as CiTO where appropriate.
• Preserve source passages for extracted claims.
• Represent versions explicitly.
• Model support and challenge as edge properties or typed relations.
• Keep model inference distinguishable from deposited facts.
• Record source independence.
• Maintain correction event history.
• Publish graph version and known coverage gaps.

Part XVIII — Applied cases

196. Case: clinical evidence

A clinical claim such as “Treatment X reduces hospitalisation in adults with condition Y” should connect to trials, population criteria, outcome definitions, effect estimates, risk-of-bias judgements and review syntheses. Challenge evidence may include failed trials, safety results or subgroup limitations.

The graph can show why the recommendation applies to one population and not another without collapsing the answer into a universal treatment verdict.

197. Case: AI model performance

“Model A outperforms Model B” depends on benchmark, split, metric, hardware, prompting, data contamination and version. A claim network should attach all of these conditions.

Leaderboards are evidence only under the tested setup. The claim should not become “Model A is better” without scope.

198. Case: education research

Education claims often depend strongly on context: age, curriculum, teacher expertise, intervention duration and measurement instruments. Evidence networks can expose why one classroom result may not generalise globally.

This is especially important when AI systems summarise “what works in education” from heterogeneous studies.

199. Case: climate and environmental evidence

Claims can connect observations, models, historical datasets and scenario assumptions. A model projection is not the same evidence type as a measured temperature series.

Provenance helps users distinguish observation, simulation, inference and scenario-dependent projection.

200. Case: humanities and historical claims

A historical claim may depend on archival documents, editions, translations and interpretive arguments rather than experiments. Evidence networks can still help by connecting claims to primary sources and documenting provenance.

The graph must respect discipline-specific standards instead of imposing biomedical support labels mechanically.

201. Case: policy evidence

A policy claim may combine empirical evidence, economic assumptions, ethical values and feasibility constraints. Evidence graphs should separate descriptive evidence from normative decision rules.

Otherwise “studies find” can be used to hide a value judgement.

202. Case: autonomous scientific agents

An autonomous agent can keep hypotheses, experiments, findings and claims as typed nodes. Each transition requires evidence: experiment executed, result observed, claim supported.

This prevents the manuscript generator from jumping directly from an attractive hypothesis to a confident conclusion.

203. Case: living knowledge bases

A knowledge base that stores claim states should refresh when source publications are corrected, when new evidence is added or when extraction models are improved.

Knowledge becomes a maintained state, not a static import.

Part XIX — The deeper model

204. Scientific papers are containers; claims are the reasoning units

The article remains the essential human-readable unit of scholarly argument. But many computational tasks operate better on claims because claims can be compared, supported, challenged and updated independently.

The future scholarly record will likely preserve both scales: documents for full reasoning and claims for machine navigation.

205. Evidence networks manufacture reversible compression

A synthesis compresses hundreds of pages into a judgement. The compression is trustworthy only if a user can reverse it: judgement → supporting claims → evidence spans → studies → data and methods where necessary.

Reversible compression is a better goal than “one truth score.”

206. Provenance is scientific memory of transformation

Science does not merely store observations. It transforms them: raw data into cleaned data, data into analyses, analyses into results, results into claims, claims into syntheses.

Provenance records the transformations so the future can audit the path.

207. Contradiction is not noise; it is structure

Disagreement identifies where models, measurements or populations differ. Evidence networks make those boundaries visible.

A healthy scientific graph should be able to represent unresolved conflict without forcing consensus.

208. Trust comes from inspectability, not from branding

A prestigious journal, large model or polished database can still contain errors. Trustworthy systems expose source, method, provenance, version and correction route.

Authority matters, but inspectability makes authority accountable.

209. The claim network is a self-correction machine

When new evidence enters, affected claims can be identified. When a source is retracted, dependent claims can be refreshed. When a claim is narrowed, the older version remains visible.

This turns scientific correction from document-by-document memory into computable maintenance.

210. The claim network is also a teaching machine

Students can learn not only conclusions but why those conclusions are held: which evidence supports them, where uncertainty remains and what would change the answer.

Evidence literacy becomes navigable.

211. AI makes claim-level provenance urgent

When humans wrote most summaries manually, readers could often inspect references and prose to reconstruct reasoning. AI can now generate millions of summaries faster than humans can audit them.

The infrastructure response should be stronger provenance, not merely more fluent generation.

212. AI should inherit the burden of proof, not escape it

A model that makes a scientific claim should carry at least the evidence discipline expected of a careful human writer: relevant sources, accurate attribution, current publication state, acknowledged uncertainty and correction.

Automation should raise the standard because it scales the consequences of error.

213. Evidence networks can become infrastructure for scientific memory

Archives preserve documents. Knowledge graphs preserve relations. Evidence networks preserve the reasons scientific claims were believed, challenged or revised.

This is a deeper form of institutional memory: not only what science said, but why.

214. The future record should preserve claim history

A claim can begin as hypothesis, become supported, narrow after replication and later be superseded. These state transitions are scientifically meaningful.

Versioned claim histories could make the evolution of knowledge directly inspectable.

215. The future record should preserve evidence independence

Independent replication is scientifically different from repeated citation. Provenance can reveal whether support branches come from separate observations or common ancestry.

Evidence diversity should become a first-class graph property.

216. The future record should preserve uncertainty as data

Confidence intervals, probability distributions, quality judgements and unknown states should remain machine-readable rather than disappearing inside prose.

A scientific knowledge system that stores only point claims loses the shape of uncertainty.

217. The future record should preserve disagreement without forcing one winner

Some questions remain contested for years. A graph can present major claim families and the evidence supporting each.

This is more scientifically honest than one synthetic answer that conceals disagreement.

218. The future record should preserve the machine’s role

If an AI extracted, normalised, classified or synthesised a claim, that activity belongs in provenance.

Future researchers should know whether a relation was asserted by an author, curated by an expert or inferred by a model.

219. The evidence-return rule: every consequential claim must return to evidence

CLAIM
  → exact scope
  → exact source state
  → evidence span
  → method / data provenance
  → supporting and challenging relations
  → appraisal
  → synthesis rule
  → dated judgement
  → correction path

If the return path breaks, confidence should fall. No amount of citation count, model fluency or graph complexity compensates for missing evidence lineage.

220. Why scientific claims and evidence networks matter to civilisation

Civilisation increasingly acts on scientific summaries rather than reading primary literature: clinicians use guidelines, governments use assessments, engineers use standards, teachers use reviews, journalists use expert syntheses and AI systems produce instant answers. Every layer of compression creates distance between decision and original evidence.

Scientific claim and evidence networks can reduce that distance. They allow a conclusion to carry its lineage: which evidence supports it, what challenges it, how it was produced, which version is current and where uncertainty remains. The purpose is not to mechanise truth. It is to mechanise the ability to return from a claim to the reasons for believing it.

That return path is one of the foundations of a self-correcting civilisation. When evidence changes, claims can change. When claims change, summaries and decisions can be refreshed. When an AI system speaks, the user can ask not only “Does this sound right?” but “Show me the evidence network that makes this claim defensible.”

Current authority routes

• Nanopublications: FAIR data containers for scientific results
• Nanopublication Guidelines
• W3C PROV-O: The PROV Ontology
• W3C PROV Model Primer
• CiTO: Citation Typing Ontology
• CiTO current version documentation
• Crossref: The Research Nexus
• Crossref: Metadata Principles and Practices
• Provenance-driven nanopublications: representing source lineage and trust networks for multi-source assertions
• Rosetta Statements: simplifying FAIR knowledge graph construction
• SciTrue: Evidence-Grounded Claim Verification in Science
• Claim Verification in the Age of Large Language Models: 2026 Survey
• SciFact scientific claim verification dataset and models
• EVI: Evidence Graph Ontology, version 1.6

Continue the Research Infrastructure series

• How Scholarly Knowledge Graphs Work
• How Persistent Scholarly Identifiers Work
• How Scholarly Indexing Works
• How Literature Search and Research Discovery Work
• How Bibliometrics and Responsible Research Assessment Work
• How Citations, References and Scholarly Linking Work
• How Systematic Reviews and Evidence Synthesis Work
• How Research Integrity and Publication Ethics Work
• Wintour House | eduKate Publishing

Publication safeguards: claim-scope, evidence, provenance, identity, source-state, contradiction, model-limit, verification, correction, freshness and archive gates.

World Return: The next time a paper, report or AI answer makes an important scientific claim, do not stop at the citation. Ask for the exact proposition, the evidence span, the method, the provenance, the supporting and challenging sources, the publication state and the date of verification. Science becomes more reliable when every important claim can carry its reasons with it.

Research route: Return to the Research Collections Directory for connected guides on evidence, methods, discovery and scholarly systems.