A verdict you can reproduce.
This page documents exactly what CitePure does to a reference, which data it trusts, how the score is computed, and where the method stops. If a result cannot be explained from this page plus the process log, that is a bug.
Five deterministic steps
Shape detection
The text is classified as a standalone reference list or a full manuscript draft by the share of lines that look like citations. A draft must contain a References, Bibliography or 参考文献 heading; the trailing reference section is then isolated.
Entry splitting and extraction
Entries are split on numbered markers, years, DOIs or blank-line groups. For each one we extract DOI, URL, year, title, author surnames and journal. Three title rules are tried in order: GB/T 7714 [J] pattern, APA "first sentence after (year)", then a longest-clause fallback. The rule that fired is printed in the report.
Two public indexes
A DOI is resolved against Crossref first and OpenAlex as fallback. A title search runs in both. Requests are throttled by 180 ms between entries to stay within polite-use limits. Both services are keyless and CORS-enabled, so the browser calls them directly.
Re-rank, then decide
API relevance order is discarded. Every candidate is re-scored on title similarity, year agreement, surname overlap and journal match. Three states are returned: Verified, Mismatch, Not found, each with a specific written reason.
Evidence, locally generated
Each row carries the original entry, parsed fields, the matched record, the DOI, the source index, the score and the reason. HTML, CSV, print-to-PDF and a bilingual Methods statement are generated in the browser.
Marker reconciliation
When a full draft is pasted, in-text markers such as [1] are compared with the reference list. Out-of-range markers and entries never cited in the body are reported separately. This checks numbering only.
Why these two indexes, and not the others
| Source | Role | Why | Status in v1 |
|---|---|---|---|
| Crossref | DOI authority | A registered DOI resolves to the bibliographic record its publisher deposited. This is the strongest available signal that an identifier belongs to a specific paper. | Live, browser-direct |
| OpenAlex | Breadth and fallback | Broader coverage including preprints and arXiv DOIs; catches records Crossref does not expose and provides a second opinion on title searches. | Live, browser-direct |
| PubMed, Semantic Scholar, arXiv API | Future redundancy | Useful, but each adds rate limits or CORS constraints. PubMed E-utilities and arXiv were tested; arXiv returns XML with no CORS header, so it cannot be called from a static page. | Not connected |
| Google Scholar | Not used | No public keyless API; scraping it would require a paid SERP layer and would make the verdict dependent on a commercial intermediary. | Not connected |
| CiNii / J-STAGE / JaLC, SciELO, Dialnet, CNKI, Wanfang | Local-language coverage | The real coverage gap for non-English bibliographies. These are the next capability layer, not a v1 promise. | Roadmap |
All lookups are read-only public metadata queries. CitePure sends only search terms (a DOI or a title) — never your manuscript text.
The exact score
Normalisation lowercases, strips diacritics, removes punctuation and a small stop-word list, and tokenises CJK text into character bigrams so Chinese titles compare meaningfully.
+ 15 × yearScore
+ 10 × surnameOverlap
+ 5 × journalMatch
− 20 (when the year differs by 4 or more)
Verified
Total score must reach 72, with title similarity ≥ 0.60 and year score ≥ 0.60 as separate gates. A high total cannot paper over a wrong title.
Title gate
Below this, a same-titled but unrelated record is not allowed to rescue a wrong year or a misattached DOI.
Candidate floor
Candidates below 0.22 title similarity are dropped as noise; if none remain, the verdict is Not found.
Authority first
If a DOI resolves and its title reasonably matches, that record outranks every title-search candidate. Otherwise a lookalike paper could "fix" a wrong year.
Thresholds are constants in assets/app.43ddfd7e.js (CFG). They are deliberately not tuned by a model. The offline self-test on the checker page asserts 16 properties of the parser, splitter, scorer and reconciliation logic.
What each result actually means
| Verdict | Meaning | Typical reasons shown |
|---|---|---|
| Verified | A record was found and its title and year agree with the submitted entry under the scoring gates above. | DOI/title resolvable; title and year match. A note is added if a supplied DOI failed but the title matched. |
| Mismatch | A record exists but the submitted entry describes something else. | Real DOI resolves to a different paper (chimera); DOI not found but a close title match exists; title mismatch; year mismatch; general field differences. |
| Not found | No comparable record in Crossref or OpenAlex for this run. | No comparable record found in either index. This is a retrieval result, not an accusation. |
Not found ≠ fabricated. Books, theses, conference abstracts, national-language journals and very recent preprints may simply be absent from these two indexes. The report never claims a reference was invented.
What this method cannot do
It does not judge scholarship
Whether a source actually supports the sentence citing it requires reading the source. CitePure checks identifier and field consistency only, and says so in every exported report.
It is not an AI or plagiarism detector
No prose style analysis, no similarity index, no comparison against a manuscript database. It also does not format citations or store a personal library.
Rule-based parsing has edges
Unusual styles, multi-line titles and references without a year can be split or parsed incorrectly. The process log exposes what was extracted so a human can spot it; the report is an aid to review, not a replacement for it.
Coverage is not completeness
Two public indexes cannot represent all scholarly output. Local-language databases are on the roadmap. A clean result is evidence about these indexes, not a certificate of universal truth.
Documented & Maintained by Julian Vance
Last updated: November 9, 2026The CitePure scoring formulas are completely deterministic, publicly auditable, and unmediated by generative models. Inquiries & reproducible methods: [email protected] or [email protected]. Read our full founder story.