Methodology
A result should show what it knows and what it does not.
SolveUnit Scholar separates explanation from reference identity checking. The reference checker uses deterministic field comparisons around external scholarly records; the explanation workflow reorganizes supplied material and must not invent missing evidence.
Reviewed by SolveUnit Scholar Engineering and Research on .
The operating principles
- Use identifiers before fuzzy text matching.
- Require multiple metadata fields to agree before verification.
- Keep incomplete, possible, mismatched and retracted results distinct.
How a citation status is produced
1. Parse the submitted reference
The checker extracts a DOI or arXiv identifier when present, then identifies a likely title, publication year and venue without rewriting the submitted citation.
2. Query independent records
Crossref, OpenAlex, Semantic Scholar, PubMed and Europe PMC are queried for each reference. A source that times out or rejects a request is recorded as unavailable, not as an empty result.
3. Merge records without hiding provenance
Records are grouped by normalized DOI or exact normalized title. Every database that supports the selected record remains visible in the result.
4. Compare important fields
An exact DOI is compared with the submitted title, year, first author and venue. Without a DOI, title coverage carries most of the score, followed by author, year, venue and cross-database agreement.
5. Apply explicit status rules
A non-DOI match needs at least 80 confidence and 0.88 title-token coverage to be verified. Scores below that remain possible matches. No confident record becomes not found only when at least three databases completed; otherwise it is incomplete.
6. Let warnings override confidence
A title, year, author or venue conflict produces metadata mismatch. A retraction signal from a scholarly database produces retracted even when the identity match is otherwise strong.
7. Return evidence for human review
The result shows checked sources, agreeing sources, field differences and the linked record. Verification confirms identity, not study quality or claim support.
Five scholarly databases
Crossref
DOI resolution and publisher-deposited journal metadata.
OpenAlex
Broad scholarly graph coverage, including many preprints and non-biomedical works.
Semantic Scholar
Independent paper identity, authors, venue and external identifiers.
PubMed
Biomedical identifiers and publication metadata from NCBI.
Europe PMC
Life-science records, open-access links and retraction indicators.
Cases that protect the status logic
These deterministic cases are covered by the reference-verification test suite. No unmeasured accuracy rate is claimed.
Verified · confidence 100
Input: Correct AlphaFold title, year and DOI.
The exact DOI and all submitted fields agree with records returned by the test fixtures.
Metadata mismatch · confidence 55
Input: Fabricated Mars title attached to the AlphaFold DOI.
A real identifier cannot rescue a conflicting title. The automated test asserts the mismatch and returns the real record.
Possible match
Input: Accurate protein structure prediction, with no DOI.
The shortened title is plausible but remains below the threshold required for a verified identity.
Incomplete
Input: No record while four databases are unavailable.
The automated test protects against describing a database outage as a missing or fabricated reference.
What this check cannot prove
- Thresholds are designed to prevent false verification, not to measure paper quality.
- Metadata providers can disagree about online-first dates, author names and publication versions.
- Reference identity must be followed by reading the source and checking whether it supports the surrounding claim.
Frequently asked questions
Is the confidence number a probability that a paper is real?
No. It is a deterministic metadata-agreement score used to rank candidate records and apply status rules.
Why can a valid DOI produce metadata mismatch?
The DOI may identify a real record while the title, year, author or venue in the submitted citation describes something else.
Why distinguish incomplete from not found?
A missing result is meaningful only when enough independent sources were actually available. Database failures must remain visible.
Are uploaded documents retained?
No document text is stored in the SolveUnit Scholar D1 application database after the requested processing.
Primary sources
- Crossref REST API documentation
- OpenAlex API documentation
- Semantic Scholar Academic Graph API
- NCBI PubMed developer resources
- Europe PMC REST API
Limits matter
Verified means that citation identity agrees with a record. It does not mean that a study is correct, relevant or appropriate to cite.