One author, two spellings, zero matches

The report looked like a disaster with a pattern in it: fourteen citations of the same author flagged as missing from the reference list, and one orphaned entry nothing ever cited. Read the two names side by side and the disaster evaporates. In the text: Al-Rashidi. In the list: Alrashidi. One scholar, one hyphen, fifteen false findings.

This case file is about the most mundane defect class we know — surname spelling variants — and why it is worth engineering care completely out of proportion to how boring it sounds. Because when name matching is done naively, it doesn't produce one small error. It produces two loud, mirrored errors at once: phantom missing citations on one side and a phantom orphan on the other, multiplied by every time the author is cited.

The shape of the failure

A citation checker pairs in-text citations with reference-list entries. To pair anything it must decide when two names are "the same", and the naive answer — the strings must be equal — fails on real documents constantly, because real documents carry names through several hands before the checker ever sees them.

in the text (14×)              in the reference list (1 entry)
(Al-Rashidi, 2019)             Alrashidi, K. (2019). Workplace
(Al-Rashidi & Osei, 2019)        adaptation in distributed teams…

naive comparison
  al-rashidi|2019  ≠  alrashidi|2019          14 "missing" citations
                                               1 "orphan" entry

folded comparison
  alrashidi|2019   =  alrashidi|2019          14 matches, 0 orphans

names invented; the defect class is real and common

Notice what the naive report does to a writer's trust. Fourteen missing citations reads as catastrophe — this is exactly the failure mode behind the checker that claimed 90% of a bibliography had failed. The writer either panics and "fixes" things that aren't broken, or correctly concludes the tool is naive and stops trusting every other row it produced — including the real ones.

Where second spellings actually come from

Nobody sets out to spell their sources two ways. The variants arrive through ordinary channels:

• Hyphens and spaces. Al-Rashidi / Al Rashidi / Alrashidi, García-López / García López. Different journals impose different house styles on the same person, and reference managers faithfully import each journal's version.
• Diacritics. Muñoz becomes Munoz the moment a reference passes through a system that strips accents — an export, a copy-paste, an older database record. The text cites one form, the imported entry carries the other.
• Transliteration. Names romanised from Arabic, Chinese, Cyrillic or Korean legitimately exist in multiple published spellings — Aleksandrov / Alexandrov — and an author's own papers may disagree with each other.
• Caps styling. Some publishers print surnames as OKONKWO, C. A checker that compares case-sensitively — or that treats the shouted form as a different token — invents a second author out of typography.
• The writer's own drift. Chapter 2 was written in March, chapter 5 in September; memory respelled the name in between and no spell-checker objects, because names aren't in the dictionary.

The author has one identity. The document has two spellings. The checker's job is to know which differences are typography and which are people.

The line we hold: fold typography, never fold people

Inside CitationLab, name comparison happens on a folded key. Before two surnames are compared, both are passed through the same mechanical normalisations: casing is levelled (which quietly de-shouts OKONKWO to Okonkwo), accents are folded so Muñoz and Munoz agree, and hyphens, apostrophes and internal spaces stop counting as differences. These folds are applied silently, and the reason they may be is important: they cannot be wrong about identity. There is no pair of distinct scholars whose names differ only by a hyphen and a capital letter appearing at the same year in the same document's own reference list — typography is not how people differ.

Everything beyond that line is treated as a difference between people until a human says otherwise. Rashid against Rashidi is one letter — and one letter is exactly how real distinct surnames differ. So the checker never silently bridges it. When the surrounding evidence is strong — same year, same co-author pattern, nothing else in the list that could claim the citation — the pair is surfaced as a suggestion, with the evidence shown, for you to confirm with a click. Confirmed, it becomes a match like any other. Declined, both findings stand. What never happens is a quiet merge the report doesn't disclose — a match you didn't make is a match you can't trust.

Why not just match harder?

The tempting fix is fuzzy matching: edit distance, phonetic codes, "close enough" thresholds. We refuse, for a reason this blog keeps returning to: a fuzzy matcher's mistakes are silent merges — two genuinely different sources presented as one, a citation credited to an entry it never meant. That is a worse failure than a false "missing", because a false missing is visible and checkable while a silent merge is invisible by construction. The checker's honest options are mechanical certainty, or a disclosed question. There is no honest middle where software guesses about identity and doesn't say so — the same reasoning that gives us the position rule for author-order traps.

There is also a structural escape from the whole problem worth knowing about: identifiers. A reference that carries a DOI can be matched by identifier, and no spelling variant on earth disturbs it — though identifiers bring their own defect class, like the DOI that resolves to the wrong paper.

The key has more parts than the surname

Everything above talked about one surname, because that is where the drama lives — but the comparison key a checker builds has more parts, and each part has its own variant problem. The year must agree, including its suffix: a document citing (Alrashidi, 2019a) and (Alrashidi, 2019b) is citing two different works, and a matcher that folds the suffix away merges two sources into one — the silent-merge failure again, one field along. Co-author separators vary by exactly the habits that produce surname variants: Al-Rashidi and Osei in prose, Al-Rashidi & Osei in parentheses, both legitimate in the same APA document, both required to match the same entry. And the presence or absence of initials — (Chen, 2021) against a list holding two different Chens — turns the surname from an answer into a question no fold can settle, which is why that case is resolved by the surrounding evidence or handed to you, never guessed.

The practical consequence: when a checker shows you a match, it should be able to show you which parts agreed and which were folded to agree. A match backed by "surname (folded), year, suffix, co-author pattern" is evidence. A match backed by "close enough" is a mood. This is the same standard we hold the report to everywhere — every number traceable to rows you can check by hand — applied at the smallest scale the system has: two names, one decision.

What a variant does to the rest of the report

One last reason this mundane class earns a case file: name variants don't stay local. A citation that fails to match doesn't just produce one wrong row — it changes the denominators everything else is computed from. The match rate drops. The missing table grows by every repeat of the citation. The orphan table gains an entry that then attracts other lookups — an unmatched entry is a candidate for every unmatched citation, so one spelling variant can seed several wrong suggestions downstream. Fixing the variant first is not cosmetic tidiness; it is draining the pool that other errors breed in. It is the reason our own workflow surfaces name-variant suggestions during review, before cross-check totals are the thing you're staring at.

What to do in your own document

1. Pick one spelling per author and let it win everywhere. The author's own most recent publications are the best authority for the romanised form.
2. Search your document for the fragments. If the surname has a hyphen, search the unhyphenated form too; if it has a diacritic, search the plain form. The variants hide from exact search — that's the point.
3. Treat mirrored findings as one finding. A "missing" citation and an "orphan" entry that share a year and a similar name are usually one spelling problem, not two defects — a healthy report reads its pairs.
4. Fix the list, not just the text. The reference-list entry is copied by every future reader; it should carry the canonical spelling.