Citations

Citations

A citation is a claim that a specific document says a specific thing. Language models produce plausible references that do not exist, and reviewers now check. The rule is simple and absolute: nothing enters the bibliography that was not resolved from a live record, and nothing is cited for a claim that was not read.

Non-negotiables

1. Resolve before you cite. Every entry comes from a record fetched now: a DOI at Crossref, a work at OpenAlex, an arXiv id, a PMID. Author names, year, title, venue, pages and DOI are copied from that record, never recalled or completed from memory.
2. Read before you attribute. Cite a paper for a finding only after reading the passage (abstract at minimum; the result section for numbers). A title is not evidence of what a paper found.
3. One canonical identifier per entry. Prefer the DOI; arXiv id for preprints without one; PMID as a secondary field. The key is firstauthorYEARkeyword and is stable once used in the text.
4. Prefer the published version over the preprint when both exist, and say so if the text relied on the preprint's numbers.
5. Never pad. A claim needs the one or two sources that establish it, not five that mention the topic.
6. Report what could not be resolved instead of leaving a best guess in the file.

Workflow

• Collect the claims that need sources, or the entries to audit.
• Resolve each to a record (references/apis.md has the exact requests).
• Read enough of each source to confirm it supports the claim as worded.
• Write the BibTeX from the record; run scripts/validate_bib.py on the file.
• Return the list of resolved, corrected and unresolved items.

Resolving. Search by title first: Crossref works?query.bibliographic=<title>&rows=3 or OpenAlex works?search=<title>&per-page=3; accept a hit only if the title matches closely and the first author agrees. With a DOI in hand, fetch https://api.crossref.org/works/<doi> for the authoritative metadata, or https://doi.org/<doi> with Accept: application/x-bibtex for a ready entry to tidy. For arXiv, http://export.arxiv.org/api/query?id_list=<id> returns title, authors and date; check Crossref for a published version of the same title before citing the preprint. For biomedical literature, PubMed E-utilities (esearch then esummary) give the PMID, DOI and journal.

Auditing an existing .bib. For every entry: does the DOI resolve, and does the resolved title match the entry's title? Do the authors and year agree? Is the venue the real one? Mark each entry verified, corrected (write the fix), or unresolved (leave the entry, flag it in the report). Duplicates under different keys are merged toward the key used in the text.

Formatting. Use @article, @inproceedings, @misc (preprints, with eprint, archivePrefix = {arXiv}, primaryClass), @book, @techreport. Protect capitals in titles with braces only where needed ({BERT}, {Transformer}). Keep doi and url fields; drop abstract and keywords. Sort by key. Let the venue's .bst or biblatex style decide the appearance; do not hand-format.

In-text use

• Cite where the claim is made, not at the end of the paragraph.
• Distinguish "X showed" (their result) from "see X" (background) from "following X" (method reused). The wording is part of the attribution.
• Numbers quoted from a source keep the source's units, precision and conditions; if the text converts them, it says so.
• When the same source appears under two names (preprint and published), cite one.

Before you hand it over

• Every \cite key exists in the .bib; every .bib entry is cited or removed.
• scripts/validate_bib.py <file>.bib reports no unresolved DOIs and no title mismatches.
• The report to the user lists unresolved items plainly. A manuscript with three honest gaps is better than one with three fabrications.