When a DOI is present
An exact DOI agreement is the strongest identity signal. Other fields are still compared so contradictory metadata remains visible instead of being hidden by the identifier.
VERIFICATION METHODOLOGY
How CiteWise turns a reference into an evidence-backed decision.
CiteWise separates text parsing, scholarly identity resolution, and field-level verification. The result is a reviewable assessment, not a claim that a confidence number is ground truth.
Last reviewed: September 9, 2026
This page describes the current production approach. Matching thresholds still require calibration against a larger labeled corpus.
PIPELINE
Six bounded stages
- 01
Preserve and normalize
The original reference remains immutable. Normalization protects identifiers while removing noise that would harm retrieval.
- 02
Extract search signals
Rules extract DOI and bibliographic fields. Parsing confidence measures extraction quality, not whether the cited work is real.
- 03
Retrieve candidates
CiteWise issues bounded queries to relevant scholarly providers using the strongest available identifiers and metadata.
- 04
Resolve one identity
Candidates are grouped and ranked as complete works. Fields from competing works are not silently combined.
- 05
Fuse field evidence
For the selected identity, title, authors, year, venue, pages, volume, issue, and identifiers receive separate evidence assessments.
- 06
Report and render
The response exposes the decision, confidence, contradictions, candidates, diagnostics, and deterministic formatted outputs.
SOURCES
Scholarly providers are evidence, not interchangeable truth
Coverage and metadata differ by discipline and provider. CiteWise records where each value came from, keeps provider failures distinct from empty results, and does not assume that the most complete record is automatically correct.
CrossrefDOI registration metadata and bibliographic recordsOfficial documentation
OpenAlexOpen scholarly works, authors, venues, and identifiersOfficial documentation
Semantic ScholarScholarly paper records and identifiersOfficial documentation
PubMedBiomedical and life-sciences citation recordsOfficial documentation
arXivPreprint metadata for arXiv recordsOfficial documentation
MATCHING
Identifiers lead; bibliographic agreement resolves the rest
When a DOI is absent
Title similarity, author overlap, publication year, venue, volume, issue, and pages contribute according to availability and reliability. Missing fields are not treated as contradictions.
When candidates disagree
Contradictions reduce confidence. Near-tied candidates produce an ambiguous decision rather than an arbitrary winner.
When a provider fails
Timeouts, rate limits, and upstream errors are diagnostic states. They do not become a false “not found” result.
DECISIONS
What each verification outcome means
| Decision | Interpretation |
|---|---|
| Verified | The leading work has strong, consistent evidence and no material unresolved contradiction. |
| Probably verified | The leading work is convincing, but the evidence is not strong enough for the highest decision. |
| Ambiguous | Two or more plausible works remain too close to choose safely. |
| Conflicting metadata | A likely identity exists, but important fields disagree across the input or evidence. |
| Not found | Available providers returned no sufficiently plausible work. This does not prove the work does not exist. |
| Insufficient evidence | The input or available provider data is too limited for a defensible decision. |
| Possible hallucination | The reference contains enough detail to search, but no coherent scholarly identity can be supported. |
AI BOUNDARY
A language model may suggest a search, never establish a fact
Deterministic extraction and provider retrieval run first. When those signals are incomplete, an optional model can propose bounded missing-field or search hints. Suggested values remain labeled as suggestions and require provider evidence before they can support verification. The model is not treated as a bibliographic authority.
LIMITATIONS
A careful result can still be incomplete
- Coverage: a legitimate work may be absent from the configured providers, especially for local publications, books, chapters, theses, or older material.
- Source quality: provider records can contain errors, delayed corrections, transliteration differences, or incomplete author and page data.
- Thresholds: current matching weights and decision thresholds are engineering defaults, not universal scholarly standards.
- Human responsibility: CiteWise supports research and editorial quality control; it does not replace checking the original publication or exercising scholarly judgment.
Inspect a result yourself.
Start with 10 provider-backed checks each month.