Citation Discipline

The one rule

You must never write a BibTeX entry from memory.

Not a plausible one. Not one you are confident about. Not "just the key fields and we will fix it later". Every entry comes from an academic API, or it does not go in the file.

Why this rule is absolute

Language models produce bibliography entries that are almost right. The authors are real, the venue is plausible, the year is close, and the paper does not exist. Audits of agent-written papers have found hallucinated reference rates above 20 percent.

Observed directly: an agent asked to fix missing references wrote nine entries in one edit. Four described no real paper. The draft compiled, the References section looked complete, and nothing flagged it. LaTeX cannot tell you a citation is fictional. Only an academic database can.

How to add a citation

1. Search for the work by title in an academic API: Semantic Scholar, DBLP, CrossRef, or arXiv.
2. Compare the returned title against what you searched for. Require a close match, not a thematic one. "Roughly about the same topic" is how the wrong paper gets cited.
3. Fetch the BibTeX from the API, not from your own text generation.
4. If nothing matches, you have three options and inventing is not among them:
   • cite a different, real work that makes the same point
   • drop the claim that needed the citation
   • keep the claim and mark it explicitly as unsupported

Fixing "?" in a compiled PDF

A ? means the text cites a key that is not in references.bib. Repair in this order:

1. Recompile with the full chain (pdflatex, bibtex, pdflatex, pdflatex). Roughly half of these are stale-aux artifacts, not missing entries.
2. For each key still unresolved, search the academic APIs using the key itself as a hint. A key like pedregosa2011scikit decomposes into "pedregosa 2011 scikit", which is usually enough to find the real entry. Re-key the fetched BibTeX to the key the text already uses so existing \cite commands resolve.
3. Whatever is still unresolved, remove the \cite command so the sentence stands without it.

Shipping a PDF containing ? is never acceptable. It tells the reader the bibliography was never checked.

When the network is down

If every lookup fails, do not conclude that every reference is fake and delete the bibliography. Distinguish "the API said no" from "the API did not answer". If nothing at all verified in a pass, assume the network and change nothing.

This is not hypothetical. A platform env once lost its CA certificate bundle after an unrelated change, so every academic lookup returned empty. Without this guard, an automated cleanup would have deleted every reference in every paper it touched.

Verify before shipping

Before the paper is done, take each entry in references.bib, search its title, and confirm it resolves. Report the count. "23 of 23 references verified against Semantic Scholar and DBLP" is a sentence a reviewer can trust. Silence is not.