Executive summary: how to match messy references correctly — and avoid falsely accusing real citations of being fabricated

A modern reference-verification system has to solve two different problems that are often accidentally treated as one. First, it must answer “which real scholarly record does this messy reference mean?” Only after that has been attempted thoroughly should it answer “does the cited work appear not to exist?” The literature reviewed here shows why this ordering matters: real citations routinely contain misspellings, missing fields, OCR damage, wrong years, incomplete author lists, incorrect identifiers and version differences, while bibliographic databases themselves can contain missing or conflicting metadata. A detector that equates “my matcher failed” with “the citation is fabricated” will therefore create false accusations. Crossref’s own real-reference benchmark is a particularly clear demonstration: its stronger search-plus-validation matcher reached an F1 of 0.9629 on 2,000 messy references, yet still produced both missed true matches and incorrect matches. [1]

The datasets in your list are not interchangeable. Crossref’s 2,000-reference set and EXmatcher are direct record-resolution benchmarks; Cora is primarily citation coreference; RenoBench evaluates citation parsing; CiteTracer and CiteAudit contain controlled mutations or fabricated-reference cases; HalluCitation Matters, GhostCite, Topaz et al., Zhao et al. and Phantom References are primarily prevalence/audit studies; and the July 2026 Manual Reference Verification Dataset is a small human-verified detector-evaluation set. Evaluating a new matcher on only one category can give a seriously misleading impression of real-world quality. [3]

Your proposed author-certified correction stratum is therefore genuinely different from most existing resources: instead of asking an annotator to infer what a damaged reference was intended to cite, the correction can provide direct evidence from an author that the original erroneous string and corrected record refer to the same intended work. That is especially valuable for measuring the failure mode that fabricated-citation detectors struggle with: a real source that looks fake because its citation has been mangled. It should, however, complement rather than replace manually verified null/fabricated cases, because a correction corpus alone will be strongly enriched for references that do have a recoverable intended target.

The problem in beginner language

Imagine that a paper contains the reference below:

Smth J, Lee A. 2021. Deep lerning for cite matchng. J Inform Retriev. 15:44-57. doi:10.1234/WRONG-ID

Smith, John; Lee, Alice. 2020. Deep Learning for Citation Matching. Journal of Information Retrieval, 15, 44–57. DOI: 10.5678/correct-doi

A simplistic detector can make a disastrous mistake. It searches the exact title, gets no result, follows the supplied DOI to an unrelated paper and reports “fabricated citation”. A good matcher interprets the evidence differently: “the DOI conflicts with the title; the title has plausible spelling/OCR corruption; the authors roughly agree; the year differs by one; let me retrieve candidate publications using several independent clues.”

A step-by-step correction

Step A — preserve the raw string. Never throw away the original citation after parsing. PDF/OCR parsers can split authors, titles or page numbers incorrectly; HalluCitation Matters explicitly found enough extraction problems that the authors used MinerU to obtain bibliography blocks and then passed those blocks to GROBID rather than relying on a single extraction route. [4]

Step B — treat identifiers as strong evidence, not infallible truth. A DOI that resolves to a record whose title and authors strongly conflict with the citation is evidence of an identifier error, not proof that the whole citation is imaginary. Both the 2026 Manual Reference Verification Dataset and modern citation-auditing taxonomies explicitly recognise “DOI points to another publication” as a distinct bibliographic failure. [5]

Step C — retrieve candidates redundantly. Search the corrected/normalised title, character n-grams, author names and combinations of fields. Crossref’s strongest 2018 approach retrieved candidates using bibliographic search and then applied a separate validation stage; EXmatcher similarly found that combining the raw reference string with parsed segment features performed better than relying on one representation alone. [6]

Step D — compare fields independently. A title may be excellent while the year is wrong. The DOI may be wrong while the title and authors are right. Modern systems such as CiteTracer explicitly classify title, author, venue, year, identifier and peripheral-field errors separately rather than reducing everything to one string-similarity value. [7]

Step E — understand versions. A preprint can later become a proceedings or journal paper with a new year, venue, page range and sometimes a changed title. Phantom References deliberately excludes ordinary venue/year/publication-status drift from its hallucination definition for precisely this reason. [8]

Step F — abstain when evidence is insufficient. “I cannot verify this automatically” is not the same statement as “this is fabricated”. RefChecker explicitly keeps unverified, uncertain and hallucination outcomes separate, which is a much safer operational model. [8]

 flowchart LR A[Raw reference] --> B[Parse but preserve raw text] B --> C{Valid identifier?} C -->|Yes and metadata agrees| D[High-confidence match] C -->|Missing or conflicting| E[Multi-index candidate retrieval] E --> F[Lexical + character + dense retrieval] F --> G[Field-aware reranking] G --> H{Calibrated confidence} H -->|High| D H -->|Middle| I[Abstain / manual review] H -->|Low| J[Escalate to more databases and web evidence] J --> K{Independent evidence found?} K -->|Yes| D K -->|No| L[Unsupported candidate] L --> M{Strict fabrication criteria met?} M -->|No| I M -->|Yes| N[Likely fabricated] 

Mermaid flowchart for WordPress: render the block with a Mermaid-compatible plug-in or block. The important conceptual rule is that “likely fabricated” is the final branch, not the first response to a failed lookup.

Existing datasets and evaluation resources

The most useful way to compare these datasets is by asking what question the ground truth actually answers. A parser benchmark can tell you whether an author surname was extracted correctly; it cannot, by itself, tell you whether a noisy reference resolved to the correct DOI. Likewise, a collection containing only fabricated references is useful for measuring recall on fabrications but cannot tell you how often the detector falsely accuses valid references.

Crossref real-reference benchmark

Crossref reference-matching evaluation framework and Marple

EXmatcher gold standard

Cora citation matching/coreference

CiteTracer benchmark

HalluCitation Matters appendix

Manual Reference Verification Dataset

RenoBench

Dataset comparison

Dataset sizes and ground-truth descriptions in this comparison follow the corresponding primary resource descriptions above. [21]

Reference-error and reference-matching research

Logan et al. — Reference accuracy of primary studies published in peer-reviewed scholarly journals

Broadus — An Investigation of the Validity of Bibliographic Citations

de Lacey, Record and Wade — How accurate are quotations and references in medical journals?

Scheinfeld et al. — Assessing Formatting Accuracy of APA Style References

Tkaczyk et al. — Machine Learning vs. Rules and Out-of-the-Box vs. Retrained

Fedoryszak et al. — Methodology for Evaluating Citation Parsing and Matching

Crossref — Reference Matching: For Real This Time

The Crossref work has already been described in the dataset section because it is both a benchmark and a matching study. The key result remains one of the most relevant empirical baselines in this whole literature: SBMV F1 0.9629 versus 0.9251 for the legacy approach on real unstructured strings, with the main gain coming from improved recall. [1]

EXmatcher

Likewise, EXmatcher is both a paper and a released gold standard. The central modelling result is that the best system combined evidence from the raw reference string and parsed segments; its reported one-match SVM experiment reached F1 0.948 before full-pipeline blocking effects and 0.941 at pipeline level. [12]

Sefid et al. — Cleaning Noisy and Heterogeneous Metadata for Record Linking Across Scholarly Big Datasets

LLM-era fabricated-citation research

The LLM-era studies are easy to misread because they use different definitions. Some count only completely non-existent works. Some count severe author mismatch. Some treat a wrong DOI or year as a hallucination. Some report candidate matches before manual checking; others manually review every flagged case. Comparing prevalence percentages without reading these definitions is therefore unsafe.

Topaz et al. — Fabricated citations: an audit across 2.5 million biomedical papers

Zhao et al. — LLM hallucinations in the wild

Sakai et al. — HalluCitation Matters

As described above, HalluCitation Matters used a deliberately high-recall screening stage followed by manual verification. The authors found 295 affected papers overall, with the proportion rising from 0.28% in 2024 to 2.59% in 2025. Crucially, they explicitly report that single-candidate cases frequently arise from OCR, parsing and fuzzy-matching noise. [4]

That makes HalluCitation Matters one of the strongest empirical motivations for your proposed study: the detector itself produces a candidate pool containing matcher failures, and manual verification must rescue real citations from that pool.

GhostCite

CiteAudit

Phantom References

Badalova and Mayr — What Current Tools Can and Cannot Do

Glynn — Analysis of the Academ-AI Dataset

What all of this evidence means

The literature is actually studying several different tasks

The reviewed work makes clear that errors upstream propagate downstream. HalluCitation Matters reports OCR/parsing and fuzzy-title failures among its candidate false alarms; Crossref reports errors from noisy strings and incomplete target metadata; Phantom References explicitly acknowledges database and extraction error as sources of false positives; and the July 2026 manual audit is designed specifically around tools incorrectly flagging genuine references. [42]

Why “not found” is not a scientific ground truth

A bibliographic database is an observation system, not the universe of scholarship. Records may be missing, duplicated, partially populated, transliterated differently or separated into preprint and published versions. Even Crossref’s own benchmark identifies incomplete/incorrect target metadata as a cause of matcher failure. [1]

This means the following inference is invalid:

Crossref returned nothing ↓ therefore the reference is not real

A more defensible inference is:

Crossref returned nothing ↓ search additional structured sources ↓ try spelling-robust and author-aware retrieval ↓ check identifiers and version relationships ↓ perform evidence-constrained web search if necessary ↓ if still unresolved, output "unverified" ↓ only call "likely fabricated" under a stricter evidence rule

The strongest recent tools increasingly follow this distinction. RefChecker has an explicit unverified state; CiteTracer has a Potential class; GhostCite manually reviews flagged references; and HalluCitation Matters manually verifies candidates rather than automatically treating its title-match threshold as final truth. [43]

Why author-certified corrections are interesting

Human adjudicators usually work backwards: they see a damaged reference and try to infer what the author meant. An author-certified correction provides a different form of evidence:

original published reference + author's correction / corrected reference ↓ certified intended identity ↓ real positive pair: damaged string → actual scholarly object

This is unusually useful for testing false-fabrication behaviour. Consider a detector that says 95% of fabricated citations are caught. That sounds excellent. But suppose it also labels 3% of heavily mangled genuine citations “fabricated”. In a real journal where genuinely fabricated citations are rare and genuine citations are abundant, those false alarms can dominate the review queue. This is the classic reason precision and base rates matter more than recall alone.

A timeline of the methodological shift

 timeline title From citation errors to large-scale hallucination audits 1983 : Broadus studies propagation of erroneous bibliographic citations 1985 : de Lacey, Record and Wade audit quotations and references 2013 : Fedoryszak et al. formalise parsing/matching evaluation 2018 : Tkaczyk et al. compare parsers : Crossref evaluates 2,000 real unstructured references 2019 : EXmatcher releases manually checked matching gold standard : Sefid et al. combine BM25 blocking, metadata and citation evidence 2024 : Logan et al. synthesise reference-accuracy literature : Glynn analyses Academ-AI 2025 : APA formatting-accuracy scoping review 2026 : Topaz et al. biomedical audit : Zhao et al. 111M-reference unmatched-citation study : HalluCitation Matters : GhostCite : CiteAudit : CiteTracer : Phantom References / RefChecker : RenoBench : Badalova-Mayr manual verification dataset 

A rigorous evaluation protocol for a 2026 matcher

The model architecture I would build

Canonical knowledge layer

Build a periodically refreshed canonical record graph from complementary sources rather than choosing one “master database”. Crossref is essential for DOI metadata; arXiv and ACL Anthology cover important preprint/NLP cases; DBLP is valuable in computer science; PubMed/Europe PMC matter for biomedicine; OpenAlex provides broad scholarly graph coverage; DataCite is important for non-Crossref DOI objects. Modern auditors already obtain robustness by combining several such sources. [44]

Represent version families: preprint, conference paper, accepted manuscript, journal version, chapter and parent book/proceedings should not be treated as unrelated records. Crossref’s observed “book instead of chapter” error and Phantom’s deliberate tolerance of publication-version drift show why this matters. [45]

Reference understanding layer

Keep two representations simultaneously:

Do not force all subsequent matching through parsed fields. EXmatcher’s best results came from combining string and segment evidence, while Crossref’s successful search-based strategy was designed to survive cases where the parser lacked enough clean structure. [46]

Candidate generation layer

Use several retrievers in parallel and fuse their candidate lists:

Measure candidate recall@k separately. If the true paper is absent from the top candidate pool, no reranker can repair the error.

Field-aware reranker

For each candidate, generate interpretable compatibility evidence:

{ "title_similarity": ..., "title_character_similarity": ..., "author_overlap": ..., "author_order_consistency": ..., "year_delta": ..., "venue_compatibility": ..., "page_overlap": ..., "doi_consistency": ..., "version_family_match": ..., "source_agreement_count": ... }

Feed these features plus text pairs into a cross-encoder or comparable learned record-linkage model. Train with hard negatives such as:

same first author + different paper same/similar title + different year conference version vs journal version book vs chapter correct title + deliberately wrong DOI paper A title + paper B authors near-duplicate proceedings records

These are much more useful negatives than randomly sampled unrelated papers because the real danger is not failing to distinguish “quantum physics” from “Victorian poetry”; it is choosing the wrong one of two very similar scholarly records.

Calibration and abstention

The production output should not be binary. Use at least:

MATCHED_HIGH_CONFIDENCE MATCHED_LOW_CONFIDENCE AMBIGUOUS_MULTIPLE_CANDIDATES UNVERIFIED LIKELY_FABRICATED

For a high-stakes research-integrity system, calibrate thresholds on held-out real data. A model with slightly lower raw recall but a trustworthy abstention region can be far safer than a nominally higher-accuracy system that confidently labels difficult valid references as fabricated. RefChecker’s multi-status design and CiteTracer’s Potential class already move in this direction. [48]

Evidence escalation

Only unresolved citations should invoke expensive search/agent stages. Check additional scholarly indices, publisher pages and repository records, then use web-search agents to find evidence. Crucially, the agent should return source evidence, not merely a yes/no judgement. GhostCite and RefChecker both use escalation after structured-source matching rather than using an LLM as the sole authority. [49]

Datasets to use in the experiment

A convincing 2026 paper should report separate results on several strata rather than merge them into one score:

These sizes and task definitions come from the corresponding primary resource descriptions reviewed above. [50]

Metrics that should be reported

Exact resolution accuracy

For references that have one known canonical target:

Exact resolution accuracy = references assigned the correct target

all resolvable references

Use exact canonical identifiers after normalising equivalent DOI URL forms. Where multiple records represent legitimate versions, publish both a strict exact-record metric and a version-family metric.

Precision, recall and F1 of produced links

This preserves comparability with Crossref:

Precision = correct returned links / all returned links Recall = correct returned links / all true links F1 = harmonic mean of precision and recall

Crossref’s SBMV provides a concrete historical baseline: precision 98.09%, recall 94.56%, F1 96.29% on its 2,000-reference benchmark. [1]

Candidate recall at k

Report whether the true item appears among the first 1, 5, 10 and other pre-registered candidate depths. This cleanly separates retrieval failure from reranker failure.

False fabrication rate on genuine references

This should be a headline metric in your paper:

False fabrication rate = genuine references labelled "fabricated"

all genuine references

Report it separately for ordinary clean references and author-certified mangled references. The latter is your core contribution.

Fabrication recall

Fabrication recall = true fabricated references correctly flagged

all adjudicated fabricated references

The 957 CiteTracer real-world fabricated cases are useful here, but because they are positive cases they must not be the only benchmark. [17]

Precision of a fabrication accusation

Fabrication precision = true fabricated among items labelled fabricated

all items labelled fabricated

This metric is ethically important. A false negative means a bad citation may escape detection; a false positive means a real author/source is being labelled as fabricated. The latter can have reputational consequences, so a production verifier should be precision-first and able to abstain.

Risk–coverage curve

Instead of publishing one arbitrary threshold, show what happens as the system becomes more selective. At 100% coverage it must decide everything; at lower coverage it can send uncertain cases to review. Plot matching error versus percentage automatically resolved.

Calibration

When the system says “0.99 probability of a match”, approximately 99% of such predictions should actually be correct. Report a calibration metric such as Brier score and/or expected calibration error, alongside reliability plots.

Per-error-type metrics

Report performance separately for title corruption, author corruption, wrong year, wrong venue, incorrect DOI, missing DOI, OCR noise, transliteration, merged references, book/chapter confusion and preprint/published-version differences. CiteTracer’s taxonomy gives a useful starting point. [7]

Operational cost

Measure database calls, latency and the proportion reaching expensive web/LLM escalation. Phantom References reports that one venue-scale audit cost roughly four US cents per paper under the paper’s May 2026 API-pricing configuration, demonstrating why escalation rather than full web search for every citation is operationally attractive. [8]

Experimental protocol

Freeze every dataset version before development. This matters particularly for fast-moving 2026 repositories and for the Manual Verification Dataset, whose released Zenodo composition differs from the initial preprint table. [19]

Split by underlying scholarly work, not merely by citation string. Otherwise two damaged variants of the same paper can leak into train and test, creating unrealistically easy evaluation.

Freeze time-sensitive bibliographic snapshots where possible. A match that was impossible because a target was absent from a database in May might become easy after metadata is added in August.

Do not tune thresholds on the test datasets. Crossref explicitly selected matcher thresholds on a separate dataset; follow the same discipline. [1]

Blind the final author-certified test set. Construct training/dev material first, lock the evaluation IDs, then reveal the test labels only once the pipeline is fixed.

Report end-to-end and component metrics. At minimum: extraction success, parsing, candidate recall, reranker accuracy, exact resolution, abstention, and fabrication classification.

Manually investigate disagreements. For every system-vs-ground-truth disagreement in a manageable test set, record whether the error came from extraction, parsing, target-database coverage, candidate retrieval, ranking, canonicalisation/version logic or bad benchmark ground truth. This error taxonomy will probably be more scientifically valuable than another aggregate accuracy score.

Baselines

The evaluation should contain at least a deliberately simple lexical baseline, a Crossref-like bibliographic search baseline, a fuzzy-title threshold resembling HalluCitation’s screening stage, a learned EXmatcher/Sefid-style record-linkage model, and currently runnable multi-source verification systems such as CiteTracer/RefChecker where licensing and APIs permit fair reproduction. Crossref legacy/SBM/SBMV results should be quoted as historical comparators on the original benchmark rather than assumed to transfer unchanged to a new corpus. [51]

What effect size should you expect?

It would be scientifically inappropriate to invent a predicted percentage improvement before running your author-certified benchmark. The existing literature gives useful observed precedents, not a guaranteed effect for your data.

On Crossref’s real strings, moving from the legacy matcher to SBMV increased F1 from 0.9251 to 0.9629, an absolute gain of 0.0378; recall rose from 0.8685 to 0.9456 while precision declined slightly from 0.9895 to 0.9809. [1]

EXmatcher’s best combined string-plus-segment representation reached pipeline F1 0.941 in its reported one-match setting, supporting the idea that heterogeneous evidence helps. [12]

CiteTracer reaches very high scores on its controlled synthetic benchmark, but those values should not be used as the expected accuracy on naturally mangled author-certified citations because the data-generating processes differ. [7]

The most defensible hypothesis to pre-register is therefore directional:

Report the actual absolute risk reduction, relative risk reduction and confidence interval after the blind test. Do not select whichever statistic looks largest after seeing the results.

Final summary and FAQ

Final summary

The literature supports a fairly coherent story once parsing, matching and fabrication detection are separated.

First, reference errors are old. Broadus, de Lacey and decades of studies synthesised by Logan show that bibliographic inaccuracies existed long before modern LLMs. A malformed reference is therefore not inherently evidence of AI fabrication. [52]

Second, matching real references is already difficult. Crossref’s carefully verified 2,000-reference benchmark demonstrates that search-plus-validation can substantially improve recall, yet even a strong F1 of 0.9629 leaves errors. EXmatcher reaches the same architectural conclusion from another direction: combine raw strings and structured fields rather than trusting one representation. [6]

Third, the LLM-era phenomenon is real enough to warrant serious checking. Topaz et al., Zhao et al., HalluCitation Matters, GhostCite and Phantom References independently report rising or clearly detectable invalid/non-existent reference problems across biomedical, preprint and conference corpora, although their definitions and adjudication methods differ substantially. [53]

Fourth, matcher failure contaminates fabrication detection. HalluCitation explicitly observes OCR/parsing/fuzzy-match noise; RefChecker warns about false positives from extraction and database coverage; Crossref documents real matching failures caused by imperfect strings and metadata; and Badalova and Mayr show that current detectors can flag many valid references. [54]

That is why an author-certified correction benchmark is useful. It targets an important stratum that positive fabricated-citation datasets do not: “we know this awful-looking reference is real, and we know what the author intended it to point to.” The strongest 2026 paper would combine that new stratum with Crossref, EXmatcher, Cora, RenoBench, CiteTracer, the Manual Verification Dataset and adjudicated fabricated cases, then evaluate the full error cascade from PDF extraction through final fabrication verdict.

FAQ: How do I build the best reference-matching model in 2026?

Build a hybrid retrieval-and-reranking system, not a monolithic LLM prompt. Preserve the raw citation, parse it into fields, resolve identifiers, retrieve candidates with BM25 + character-level matching + dense retrieval from multiple scholarly databases, fuse candidates, rerank with a model that understands title/author/year/venue/identifier evidence and publication versions, calibrate its confidence and provide an abstention state. Escalate only unresolved cases to additional databases and evidence-constrained web search. This architecture follows the strongest recurring ideas in Crossref SBMV, EXmatcher, Sefid et al., CiteTracer, GhostCite and RefChecker. [55]

For training, favour hard real errors and hard negatives. Synthetic corruptions are useful for coverage, but they should supplement—not substitute for—naturally damaged references and blind real-world evaluation.

FAQ: How do I reduce false fabrication flags?

The single most important rule is:

Failure to match is a reason to investigate, not a fabrication label.

Use a three-stage interpretation: matched → unresolved/ambiguous → likely fabricated. The fabricated label should require stronger negative evidence than the unresolved label. Search several independent sources, test alternative spellings, handle transliteration, tolerate year/venue version drift, verify whether an incorrect DOI simply points to a different real object, and look for parent/child publication relations. RefChecker’s distinction between unverified and hallucination and CiteTracer’s Potential category are useful precedents. [48]

Then optimise and publish the false fabrication rate on author-certified real references. A detector should not be considered safe merely because it catches most synthetic fabrications.

FAQ: Should I train a giant language model from scratch?

No. The evidence reviewed here gives no reason to believe that training an end-to-end foundation model from scratch is the efficient path. The hard problem is not simply language understanding; it is retrieving and reconciling external bibliographic evidence. Crossref’s search-based approach, EXmatcher’s blocking/classification pipeline, Sefid et al.’s BM25 blocking, CiteTracer’s retrieval cascade and RefChecker’s multi-source verification all benefit from explicit retrieval infrastructure. [56]

Fine-tune relatively small specialised components where labelled data justifies it, and use large models where they add unique value—noisy parsing, difficult semantic comparison or constrained evidence-search escalation.

FAQ: Should DOI matching be the first thing I do?

Yes, but do not treat it as infallible. A DOI that resolves and whose metadata agrees strongly with the citation is nearly ideal evidence. A DOI that resolves to a completely different title should be classified as an identifier conflict and trigger candidate recovery. The Manual Reference Verification Dataset explicitly treats a DOI that points to another publication as a problematic-reference mode. [18]

FAQ: Is title similarity enough?

No. HalluCitation Matters uses a 0.9 normalised-Levenshtein title threshold as a screening mechanism, but its authors then manually verify candidates and report noise arising from fuzzy matching and extraction. EXmatcher found better performance from combining string and field evidence, and Crossref’s SBMV separately validates retrieved candidates. [57]

FAQ: Which dataset should be the primary benchmark?

For existing direct resolution ground truth, start with the Crossref 2,000-reference benchmark because it most closely asks “what DOI does this real messy string mean?”. Add EXmatcher for another manually checked resolution environment. [58]

For your new contribution, make the author-certified correction corpus the main stress-test stratum and report Crossref/EXmatcher as established comparators. Add the Manual Verification Dataset to test false flags, CiteTracer to test controlled error modes and genuine fabrications, RenoBench to isolate parser quality, and Cora to test coreference robustness. [59]

FAQ: What should the paper’s main metric be?

Do not use only one metric. For matching, make exact target accuracy plus link precision/recall/F1 primary. For your research-integrity contribution, prominently report false fabrication rate on author-certified genuine references, fabrication precision, fabrication recall and coverage at high precision.

The key scientific claim should not be “our model has 98.x% accuracy”. A much more meaningful claim would be: given genuinely real references that were damaged badly enough to require author correction, how often can the model still recover the intended scholarly object, and how rarely does it falsely call that object fabricated?

FAQ: What practical implementation details are easiest to overlook?

Cache external lookups, store provenance for every candidate, preserve raw strings, normalise Unicode, test transliterated author names, distinguish missing from conflicting fields, collapse known version families, detect book/chapter and proceedings/paper relations, version every external data snapshot, and make every final fabrication verdict inspectable by a human. CiteTracer reports that most benchmark cases can be resolved in deterministic/cached stages while a minority reach expensive judging, illustrating why a cascade is operationally sensible. [7]

Finally, never expose a naked “FAKE” label without its evidence. Return something closer to:

{ "status": "UNVERIFIED", "best_candidate": "...", "candidate_score": 0.61, "conflicts": ["DOI resolves to another work"], "supporting_sources": ["Crossref", "OpenAlex"], "sources_searched_without_match": ["DBLP", "ACL Anthology"], "recommended_action": "manual_review" }

That design makes the system useful to editors and researchers while acknowledging the epistemic reality documented throughout this literature: bibliographic infrastructure is incomplete, citations are noisy, and matcher failure is not the same thing as scientific fabrication.

Reference table with links

Research status: sources were checked against primary or official pages where available, including Crossref, arXiv, ACL Anthology, GitHub/GitLab project repositories, Zenodo, Hugging Face, BMJ, SAGE, Érudit and institutional author/publisher pages. Fast-moving 2026 preprints and repositories may change after 13 August 2026; reproducible experiments should record exact dataset versions, commits and access dates. [60]