Check DOI citations and bibliography consistency

Send one to five citations to AUN to retrieve Crossref metadata, identify invalid DOI syntax, and flag low title similarity. This API is available in the free beta with usage quotas. No card or wallet is required.

Request

POST https://agentutility.network/api/bibliography
Content-Type: application/json

{"citations":[{"doi":"10.1038/nature14539","expected_title":"Deep learning"}]}

An optional Bearer credential provides an individual allowance. Each submitted citation consumes one quota unit once admitted, including failures. Shared limits can return HTTP 429; respect Retry-After. See quotas and abuse protection and data handling.

Interpret the result

• resolved: Crossref returned metadata and any supplied title did not fall below the similarity threshold. This is not a semantic accuracy verdict.
• title_mismatch: title-token similarity is below 0.8. Review the citation and the returned source.
• invalid_doi: input syntax was rejected.
• not_found_in_crossref: the record was absent from Crossref; another DOI agency may hold it.
• provider_error: lookup failed. An HTTP 200 bibliography response can contain these failures; inspect each row and delivery_status.

Title comparison uses Jaccard overlap of sets of normalized letter/number tokens. Word order and repeated words do not affect the score. Case, punctuation and Unicode compatibility forms are normalized. This does not check scientific claims, full text, authenticity or complete retraction status. Full OpenAPI contract.

First diagnostic evidence — 20 October 2026

12 of 13 author-authored offline cases produced the expected classification. This is a small diagnostic set built by AUN, not independent certification, a sampled accuracy estimate or an enterprise SLA. It uses two publisher reference titles and deliberately altered inputs. Provider absence/outage cases are simulated. No production algorithm was changed to fit these cases.

Among nine title-comparison cases, three of four deliberately incorrect titles were flagged and none of five matching variants were flagged. These counts apply only to this set; they do not predict performance on customer bibliographies.

Known failure: negation

The altered title Why Most Published Research Findings Are Not False adds a negation to the original title. The token score is 0.875, so the current API returns resolved instead of a mismatch warning. Do not treat resolved as proof that the supplied wording is correct. A semantic or stricter comparison and a broader, independently labelled dataset remain future work.

All cases, including the failure

Reproduce and inspect

Download the source and cases, extract them and run node quality/run.mjs with Node.js 24. No network calls, dependencies or credentials are needed for the offline diagnostic. Compare the SHA-256 identifiers in the full result JSON. The archive includes the current audit implementation and supporting DOI module; the diagnostic provider is an explicit fixture, not a real Crossref lookup.

A separate single live check submits four public citations through the public endpoint. It is not an uptime or latency benchmark. Its timestamp and provider provenance apply only to that run.

Reference sources

• Nature: Deep learning
• PLOS Medicine: Why Most Published Research Findings Are False

For agents and developers

Machine-readable catalog · Automatic credentials and delegated free orders · More examples. No paid plan, paid execution mandate or independent quality certification is available.