Reference Authenticity Audit

Establish both whether the work exists and whether the supplied citation identifies it correctly. A real article with a wrong author or year is a metadata mismatch. An unsuccessful search does not prove fabrication. Bibliographic verification does not establish that a paper supports the manuscript's claim.

Set the scope

Identify the supplied reference list, intended versions, and whether the task is audit or authorized correction. Preserve original entries and stable IDs. Do not rewrite manuscript prose during a reference audit. Search only the bibliographic information needed for the task, not private manuscript passages.

Create or reuse the reference ledger before retrieval. Preserve candidate groups without merging them, and share only original inputs and neutral indexing between verifiers. Reuse unchanged verified identities across later format checks.

Require two independent checks

The coordinating agent assigns the same references to two distinct agents, each with the original citation and intended version. Neither sees the other's findings or a suggested verdict before completing its own record. Splitting the list between two agents is not cross-verification. A second persona in one conversation is not a second agent. A verifier explicitly assigned a leaf role completes its own check without spawning further verifiers.

Each agent independently begins each reference on Google Scholar, preferably with its full title in quotes, then narrows or expands using author and year if necessary. It opens the matching record/version and records its actual search and inspected sources. It must check all four fields, including the complete title and author list, rather than accepting a truncated search snippet. Read field decisions for exact-match and version rules.

Each agent must independently corroborate the required fields against primary publication evidence for the intended version, such as the publisher record, official proceedings, or the article itself. For a preprint, use its authoritative repository record. Do this even when Scholar appears complete. Do not treat two agents reading the same automatically extracted metadata as independent corroboration. The agents may legitimately reach the same primary source; record that shared source dependence rather than claiming two independent underlying sources. Resolve partial fields with additional primary evidence.

Distinguish a tool/parser failure, a site access restriction, and a genuine search with no matching result. If the retrieval tool cannot read Scholar, try an already available, authorized ordinary browser interface when feasible. A neutral capability check can identify a usable route, but each agent still retrieves its own evidence. Stop at an actual access challenge or after the available normal route also fails; do not loop through endpoints or automate bulk retries. Record each attempted route and its outcome. This supplements the Scholar-first check; it does not silently replace it. Do not bypass access controls, CAPTCHA, or bulk-access limits. If either agent or the required sources are unavailable, provide useful partial findings and explicitly leave dual verification incomplete. Never invent a second review, a search result, a timestamp, or a source visit.

Scholar's official help explains citation exports, coverage limits, and automated metadata extraction. Index absence and extraction error are possible. Do not turn its search result or citation count into a quality or authenticity guarantee.

Cross-check a BibTeX record against the actual export

For every supplied or requested BibTeX entry, obtain the intended result's actual Cite → BibTeX output or a retained capture with verifiable provenance. Compare its title, ordered authors, year and venue with each verifier's independent identity evidence and primary record. Record export provenance and exact working-entry equality separately from existence, version and process. Evidence that the paper exists can resolve its identity while the BibTeX step remains pending. A working entry can match a capture while a year or version conflict still blocks acceptance. Include all in-scope entries in coverage totals. When the export is missing or conflicting, report that outcome and preserve the original. Only the captured text can supply the accepted BibTeX, including its key. Keep any independently verified supplemental metadata in the ledger.

Capture each agent's evidence

For every reference, record the original ID; exact query and observed source URLs; inspection time; intended work/version; observed full title, complete ordered authors, publication year, and venue; field matches or discrepancies; and source/access limitations. Use actually observed values. Mark unknown fields unknown. Retain a reviewable source artifact for the consequential field decisions when permitted, such as the downloaded export, a minimal metadata excerpt, or a screenshot of the relevant public record, with its source and capture time. Distinguish directly inspected content from search-index snippets. Report when evidence could not be retained; a written assertion of a visit is not execution proof. Keep account details, session parameters, and unrelated page content out of public records. Do not bulk scrape.

Treat supplied snapshots as snapshot evidence. Identify who captured them and when if known. A synthetic fixture or replay can test decision rules but cannot be reported as a live Scholar verification. Two independent analyses of one capture are not two independent source retrievals.

Reconcile without voting

Compare the two completed records field by field. Agreement is necessary for a clean dual-verification verdict but does not override conflicting primary evidence. Investigate differences in edition, version, online-first year, full author list, venue, and DOI. Retain both initial records and the resolution evidence; a third check may help but does not erase disagreement.

Use separate status dimensions rather than a vague pass:

A reference is fully verified under this workflow only when both independent Scholar-first checks and primary-source corroborations are complete and the required fields and intended version match with no unresolved evidence conflict. For BibTeX acceptance, additionally require a traceable matching Cite → BibTeX capture and exact working-entry text. Correctable mismatches remain findings until the corrected entry has been checked. Do not label a work nonexistent solely because it was not found.

Deliver a bounded record

Lead with the number and scope actually checked, then a concise table of consequential findings and kept entries. Attach the two source records and any reconciliation when a full audit is requested. Separate confirmed facts, proposed corrections, and unresolved facts. Keep exact title punctuation, spelling, technical terms, and required bibliographic formatting. The library's prose preferences do not authorize rewriting publication titles. Stop when the scoped references have explicit outcomes and actionable remaining checks.