The APA default that broke every Vancouver thesis

For a stretch of time we're not proud of, our own system knew the citation style of every document it read — worked it out correctly, wrote it down — and then threw the answer away. Every thesis that followed was treated as if it were written in APA. For most documents, nothing visibly went wrong. That's what made it dangerous.

This is a case file about the most common defect in citation software — not a crash, not a wrong calculation, but a silent assumption. It's the story of why almost every checking tool speaks author–date as its native language, what that does to the numbered styles half the scientific world writes in, and how we found and fixed the version of this bug that lived in our own code.

Every checker has a native language

Citation styles split into two great families. Author–date styles — APA, Harvard and their many cousins — put a name and a year in the text: (Nguyen, 2021). Numbered styles — Vancouver, IEEE — put a number there instead: [14], with the reference list keyed by those numbers. The two families don't just format differently; they carry their information in different places. Author–date citations are self-describing. A numbered citation means nothing without its list.

Software built to check citations has to pick a shape to expect, and the industry overwhelmingly expects author–date. Feed most tools a numbered thesis and they'll do one of three things: read the bracketed numbers as noise, read them as arithmetic, or — the worst case — read the reference list with an author–date parser, find no parenthesised years where it expects them, and conclude the list contains nothing at all. We've written up what that looks like from the student's side: a report claiming 90% of citations are missing, in a thesis whose true error count was two.

A default is a decision someone forgot they made

Nobody sits down and decides "our software shall mistreat medical theses." Defaults form quietly. A team builds a citation parser; they need test documents; the documents nearest to hand come from psychology, education, management — fields that write in author–date. The parser grows to fit what it's fed. Months later, author–date isn't one style among many anymore. It's baked into variable names, into what a "citation" fundamentally is, into every assumption downstream. The default has become invisible — which is different from being gone.

The most dangerous defaults aren't the wrong ones. They're the ones nobody remembers are defaults.

The bias has a demographic root, too. Author–date fields produce the bulk of theses that citation tools are demonstrated on, marketed to, and tested against. Numbered styles concentrate in medicine, nursing, engineering — disciplines with their own tooling traditions and, frankly, less patience for software that fails them once. So the feedback loop that would have caught the bias never quite closes.

Our confession: detected, then discarded

Here's our own version, because this post would be dishonest without it. CitationLab has long had a style detector. It reads the document's citations and reference entries and works out which family they belong to — and it was right about Vancouver theses. The defect was in what happened next: the detected style was handed to the screen, and never to the engine. The display layer knew the thesis was Vancouver. The machinery that formatted references and rendered results quietly fell back to its default. Which was APA.

Detector:  "This document is Vancouver."   → shown on screen, then dropped
Engine:    "No style specified. Assuming APA."
Result:    A Vancouver thesis, processed
           end-to-end under APA rules.

Notice the cruelty of the failure mode. For the majority of documents — the author–date ones — the default happened to match the truth, so everything looked fine. The system passed its tests, handled its demos, satisfied most of its users. Only writers in numbered styles paid, and from the outside their symptoms looked like their errors: mangled references, odd matches, formatting that wasn't what they wrote. A silent default doesn't just break things; it breaks them in a way that blames the user.

The fix was not "improve the detector." The detector was fine. The fix was architectural: the detected style became a value with one home and many readers — decided once, when the document is first read, then carried through every later stage: matching, checking, annotation, formatting, export. No stage re-guesses. No stage falls back. If the style is Vancouver at extraction, it is Vancouver at export.

What "detected" actually means

A style isn't declared anywhere in a thesis — it has to be inferred from evidence, and real documents are messy about it. A thesis is rarely 100% one style; it's more often four-fifths one thing, with a scattering of citations pasted in from supervisors' comments, older drafts, or reference managers set to the wrong output. So detection is a matter of dominance: which family does the document's own evidence overwhelmingly point to? That dominant style becomes the working assumption — disclosed, not hidden — and the stragglers get flagged against it rather than silently redefining it. (How to run that diagnosis on your own document is its own guide.)

Disclosure is the part we'd argue is non-negotiable, in any tool. A checker that never tells you what style it thinks you wrote in has decided it doesn't need your agreement about the single assumption everything else rests on. You should get to see the claim "this document is Vancouver" — because you are the one person who knows whether it's true.

How to tell if a tool is assuming

Three questions expose a silent default in any citation software, ours included:

1. Does it name your style? Somewhere visible, before the results, it should say what referencing style it read your document as. If it never says, it never asked — it assumed.
2. Do the counts survive contact with your document? A tool should tell you how many citations and how many reference entries it parsed. Compare the reference count against your own list. A numbered thesis "parsing" to a handful of entries — or none — means the reader, not the writing, failed.
3. Do the failures follow a style boundary? If everything flagged happens to be your numbered citations, or your entries without parenthesised years, the tool isn't finding your mistakes. It's finding the edge of its own assumptions.

And a note for the numbered-style writers who've been burned before: the failure isn't intrinsic to your style. Vancouver is, if anything, easier to check than author–date once it's actually read as Vancouver — the numbers make citation-to-reference links explicit. The failure was only ever in tools reading one language while you wrote in another. (If you update references in a numbered thesis, the renumbering problem is its own minefield — our sister product covers updating without breaking your numbering.)

The lesson we kept

The style bug taught us a rule we now apply everywhere: any value important enough to detect is important enough to carry. Detecting and dropping is worse than never detecting — it means the system knew better and acted worse. Style was the first value to get this treatment; the same architecture now moves every important fact from first reading to final export, with no stage allowed to fall back to a private guess. Your thesis gets read in its own language, start to finish — and your Ref[In] Score is computed against the style you actually wrote, not the one a parser found convenient.