'Nelson SD, W.' — anatomy of a mangled author

Somewhere in a review screen, a tool once presented an author named "Nelson SD, W." No human has ever been called that. It is not a typo and not an import accident — the underlying reference was flawless. It is what happens when a correctly written Vancouver author list is read by software speaking APA: the same characters, parsed under the wrong grammar, yield people who don't exist.

This case file dissects the mangling step by step — because the failure is completely deterministic, which means it is completely preventable, and because the cascade it triggers downstream is a perfect illustration of why we treat refusing to parse as a feature.

Two grammars, one string

Author-date styles and Vancouver don't just format differently — they use the comma to mean different things. APA writes each author as Surname, Initials, and separates authors with commas and an ampersand: Nelson, S. D., & Wong, P. T. The comma lives inside each author. Vancouver writes each author as Surname Initials with no periods — Nelson SD — and uses the comma between authors: Nelson SD, Wong PT, Achebe K. Same punctuation mark, opposite job.

A parser that assumes one grammar will confidently misread the other — the same root failure as the APA default that broke every Vancouver thesis, this time inside a single name.

the entry (Vancouver, correct)
  Nelson SD, Wong PT, Achebe K. Sleep and shift work…

read under APA grammar — "comma separates surname from initials"
  token 1   "Nelson SD"   → surname: Nelson SD   initials glued on
  token 2   "Wong PT"     → initials of author 1?  surname of 2?
                            resolver panics, keeps first letter: "W."
  remainder                 → one concatenated pseudo-author

result    Nelson SD, W.            two broken names, one lost

read under Vancouver grammar — "comma separates authors"
  Nelson S. D. · Wong P. T. · Achebe K.        three people

names invented; the construction is textbook Vancouver

Why a wrong parse is worse than no parse

"Nelson SD, W." is not just cosmetically ugly. Parsed names are the input to everything downstream, and garbage compounds:

• Matching fails silently. The in-text citation "[7]" or "(Nelson et al., 2022)" tries to pair with an entry whose recorded surname is "Nelson SD" — one character of difference is all it takes, and here there are three.
• Phantom people enter the ledger. "W." can now collide with a genuine Wong elsewhere in the list, or fail to — either way the author-level accounting is fiction.
• Author-count rules misfire. Et al. thresholds, first-author ordering, position checks — all of them now operate on two mangled names instead of three real ones.

A parser that guesses produces errors that look like data. A parser that refuses produces a question that looks like a question.

The asymmetry deserves stating plainly: a false refusal costs seconds and is visibly a refusal; a false parse costs hours and is invisibly a lie. Any system that touches your bibliography should be designed around which of those two mistakes it would rather make — and be able to tell you.

That is the design rule this class taught us. When our parser meets an author string that doesn't satisfy the document's grammar cleanly, it does not emit its best guess as if it were fact. The entry is held as unparsed, with a reason — the raw string preserved and shown, the explanation attached ("author list doesn't match the document's citation grammar"), the item still on the ledger and still yours to fix or confirm. A refusal costs you ten seconds of reading. A confident mangle costs you an afternoon of chasing phantom mismatches.

Grammar first, names second

The other half of the fix is ordering: the document's citation grammar is decided once, before any name is parsed. Style is detected from the document's own evidence — bracket shapes, separator habits, initials placement — and then every entry is read under that one grammar. What the parser never does is decide per-entry: a list that parses entry 14 as APA and entry 15 as Vancouver isn't flexible, it's inconsistent — the styles disagree about what a comma means, so mixing them inside one list makes the output meaningless. One document, one grammar; exceptions surface as refusals, not as silent style-switches.

What a refusal looks like on screen

Because "the parser refuses" sounds like an error page, it's worth describing what it actually is: a row like any other, holding the entry's raw text exactly as your document wrote it, a plain-language reason, and the same editing tools every row has. Nothing about the entry is lost — the title, year and identifier that did parse cleanly still work, still match, still count. Only the author list is held open, and fixing it is usually one glance: you can see it should read three Vancouver authors, you confirm or correct, and the row rejoins the fully parsed population. Compare the alternative: "Nelson SD, W." sails through looking parsed, fails to match everything, and you meet it half an hour later as four unexplained rows in a cross-check report. The refusal isn't the tool being modest. It's the tool declining to spend your trust on a guess.

The rest of the mangling family

"Nelson SD, W." is the cleanest specimen, but the same wrong-grammar mechanism produces a whole family of fictional people, and it's worth knowing their faces:

• The split compound. Van der Berg J read by a parser that treats spaces as name boundaries becomes author "Van", author "der", and a surname "Berg J". Particle surnames — van, von, de la, bin, abu — are the stress test of any name grammar, because the particles are lowercase, common, and structural.
• The swallowed suffix. Nelson SD Jr — is "Jr" a generational suffix or a two-letter initial pair? Under the wrong grammar a junior becomes a fourth initial; under a timid one he becomes a separate author. The correct read depends on the style's suffix conventions, which is one more reason grammar must be decided before names are cut.
• The hyphen harvest. García-López M, Chen W parsed with a rule that splits on hyphens yields "García" and a phantom "López M" — which then fails to match the correctly-hyphenated citation in the text, twice over.
• The consumed conjunction. In styles that write "and" between the last two authors, a parser reading Vancouver's comma grammar can absorb the word into a surname: author "and Achebe K". One token of house style, one fictional co-author.

Notice, too, where these specimens come from: almost never from the writer's own typing. They are produced by the same supply chain that delivers every other name defect — reference managers exporting under one style what they imported under another, publisher metadata written for machines and read by different machines, copy-paste across documents with different conventions. The writer is usually the last innocent party in the chain and the first person the report appears to accuse; part of writing an honest checker is making sure the row's language blames the string, not the student.

Every one of these specimens is deterministic: same input, same grammar, same wrong people, every time. That is what makes the class fixable — and what makes it testable. Each mangled shape we meet joins a regression corpus of author-list constructions, so the parse that handles "Nelson SD, Wong PT" correctly is provably not un-fixed by next month's change to particle surnames. Parsing names is not a problem you solve once; it is a problem you stop re-breaking.

Spotting mangled names in your own bibliography

1. Look for capital-letter pairs riding surnames. "Nelson SD" in a list that elsewhere writes "Nelson, S. D." means two grammars are cohabiting — usually a paste from a differently-styled source.
2. Look for one-letter authors. A bare "W." or "K." in an author list is almost always a fragment of a neighbour, not a person.
3. Count the authors against the source. The publisher's page says three authors; your entry shows two. Somewhere a comma was misread — by you, your reference manager, or a tool.
4. Keep one style per document. If you're switching styles late in the draft, do it properly — the breakage has a pattern — and renumbering hazards are their own subject, one our sister tool handles when updating references without breaking Vancouver numbering.