What is citation chaff, and why is it on my screen?

You upload a chapter, the extractor runs, and the review screen shows rows that are plainly not citations. A table caption. A postal address. Half a figure credit. This is not the tool failing. It is the tool showing you its work.

We call those rows chaff — the borrowed farming word for what comes off the field along with the grain and has to be separated out afterwards. Every citation extractor produces some. What differs between tools is whether you ever find out.

Why anything that looks like a citation is picked up

An in-text citation has no reserved syntax. It is prose, formatted by a human, in one of a dozen conventions, sometimes inconsistently within the same chapter. There is no marker that says "this is a citation and that is not".

So the extractor works on resemblance: a bracketed name and year, a superscript number, a parenthetical with a page range. And plenty of things that are not citations resemble citations very closely.

  (Okonkwo, 2019)              a citation
                               name, year, in a sentence

  (Figure 3, p. 44)            NOT a citation
                               a cross-reference to the
                               author's own figure

  (Department of Health,       a citation — an
   2020)                       organisation as author

  (see Chapter 2)              NOT a citation
                               a pointer inside the thesis

  (Adeyemi, in press)          a citation, with no year
                               to hook it on yet

Two of those four are chaff. Nothing about their shape says so — you know because you understand what the sentence is doing, and the extractor does not.

The choice every extractor makes, whether or not it admits it

Faced with that ambiguity, a tool can lean one of two ways.

It can be strict: only take what is unmistakably a citation. The review screen comes back clean, the user is impressed, and the citations it quietly declined to collect are simply absent — from the screen, from the cross-check, and from the report. Nobody notices, because a thing that was never extracted leaves no trace.

Or it can be generous: take everything that plausibly resembles one, then screen. The review screen comes back messier, the user has to look at rows that are obviously not citations, and nothing real was lost on the way in.

A missed citation is invisible. A false one is merely annoying. Given a choice between the two errors, we take the annoying one every time.

We take the second path, and the messy screen is the visible cost of it. It is a cost we are content to pay, because the alternative fails silently — and a reference checker that fails silently is worse than no reference checker, since it also supplies confidence.

What screening actually does

Screening is a pass over the collected rows that separates the citations from the rest. A row screened out is not deleted. It is set aside, visible, and reversible — because the screen is a judgement too, and judgements are sometimes wrong.

Where a row is genuinely ambiguous, it stays in front of you rather than being decided quietly in either direction. The screen is deliberately not the last word. It is a first pass that you can overrule, and overruling it takes one click.

When the screen gets it wrong — in either direction

Two failure modes, and they are not symmetrical.

A real citation screened out as chaff is the serious one. If it stayed screened, your cross-check would never look for its reference, and a genuinely missing entry would be reported as no problem at all. That is why screened rows stay visible instead of vanishing: you can read the pile and put anything back.

A non-citation left in is the mild one. It flows into the cross-check, finds no reference — because there is no reference, it is a figure caption — and appears as a missing citation. You look at it, recognise it, and dismiss it. Irritating, and self-correcting. We wrote about a report that was mostly this kind of noise in the checker said 90% of my citations failed.

Putting a row back, and what the record remembers

You can reinstate any screened row. So can the AI review, which reads the screened pile and flags anything it thinks was set aside wrongly.

Both routes are legitimate and both are recorded — but they are recorded as different events, because they are different events. The certificate names which one put each row back. "The screen flagged this and I overruled it" and "the AI review recovered this" say different things about how much attention a line received, and the person reading your report is entitled to know which they are looking at.

  REINSTATED FROM SCREENING            3 rows

  (Department of Health, 2020)   reinstated by you
  (WHO, 2021)                    reinstated by you
  (Ministry of Works, 2018)      reinstated by AI review

  the screened pile itself is not printed
  only what came back, and who brought it

Note what is absent: the raw screened pile does not appear in the certificate. Printing every row an extractor considered and rejected would produce pages of noise and imply the document was full of problems it never had. Only the reversals are reported, because only the reversals are decisions.

Reinstatement also survives a rebuild. Upload a revised chapter and the row you promoted stays promoted — the decision belongs to you, not to the copy of the document you happened to be looking at when you made it. The conservation rule behind that is in nothing gets lost.

Numbered styles produce a different pile entirely

Everything above assumes an author-date convention, where a citation carries a surname and a year and therefore carries its own evidence that it is a citation. Numbered styles — Vancouver, IEEE, and the rest — do not.

In a numbered thesis a citation is [14]. That is the whole of it: a bracket, a digit, a bracket. And a thesis is full of brackets containing digits that are not citations at all.

  [14]        a citation

  [1-3]       three citations, or a range
              in an equation

  [2]         a citation — or a footnote
              marker in a table

  [0,1]       an interval in the methods
              chapter, not a citation at all

An author-date citation announces itself; a numbered one does not. The year that would normally confirm a candidate is simply absent, so there is nothing to corroborate against and the screening has less to work with. The result is that numbered theses generate a larger screened pile, and more of it is genuinely ambiguous rather than obviously wrong.

This is worth knowing before you open the screen, because the reaction to a long pile should not be "the tool is broken". It is a property of the convention you are writing in. A methods-heavy engineering thesis in IEEE will always give a busier screen than a humanities thesis in Harvard, and neither says anything about the quality of either.

What to actually do with your screen

Practically: skim it, do not audit it. The rows that matter are the ones that look like real citations, and they stand out from figure cross-references and address fragments immediately to the person who wrote the chapter.

You are the only one who can tell, and it takes a fraction of the time it takes to read the same list of everything a stricter tool decided not to mention.

What this does not claim

Chaff is not a measure of anything about your writing. A chapter with a lot of it is usually a chapter with a lot of tables, or a numbered style, or an author who cross-references their own figures carefully — none of which is a defect. The count is a property of the document's shape, not its quality, and nothing in the report treats it as a finding of misconduct or of carelessness.

Nor is the screen a guarantee. It is a first pass over an ambiguous problem, offered to you with its workings visible, precisely because it is the kind of judgement that a person reading their own chapter can make far better than software can.