Ref[In] — naming the thing a citation check is actually for

Every tool in this space describes an action. You run a citation check. You do a reference check. You validate a bibliography. The verbs are fine, but they all share a blind spot: an action can complete successfully and leave the document no sounder than it was.

We kept running into the gap in our own build. A check would finish, report no errors, and be wrong — not because the code failed, but because "finished without errors" was never the property anyone actually wanted. What they wanted was a claim about the document: these references are sound, and here is the evidence. There was no accepted phrase for that, so we named it. REFerence [IN]tegrity — Ref[In] for short.

This post is about what the name asserts, and — just as importantly — what it refuses to assert.

A check can pass and still be wrong

Here is the failure that made the distinction concrete. A citation in the text and an entry in the bibliography agreed on surname and year. The matcher paired them and reported a clean result.

In text      …reduced readmission in the intervention arm (Okafor, 2021)…

Bibliography Okafor, T. (2021). Retinal imaging in screening
             programmes. Journal of Clinical Ophthalmology.

Verdict      MATCHED — surname ✓  year ✓

The thesis was about hospital logistics. The reference was an ophthalmology paper by a different Okafor entirely. Surname plus year is not identity; it is a coincidence that happens constantly in a discipline where a few surnames and a twenty-year window cover most of a literature. The check completed. The pairing was wrong. Nothing in the vocabulary of "did the check pass?" could express the difference.

A tool that only reports whether it finished is answering a question about itself. Integrity is a question about your document.

The same shape shows up elsewhere. A numbered citation like (12) carries no author and no year at all — nothing intrinsic to it identifies what it points at, so its soundness is entirely a property of the list it refers into. A bibliography can balance perfectly while two of its entries are the same source under different formatting. In each case the process succeeds and the document is not sound.

Why "integrity", of all words

It is a loaded choice and we did not make it casually. In ordinary speech integrity is a virtue — a property of a person, the opposite of dishonesty. That is not the sense meant here, and if it were, the name would be an accusation waiting to happen.

The sense we mean is the engineering one, and it is older and more useful. A bridge has structural integrity: its members carry the loads they claim to carry, or they do not. A database has referential integrity: every foreign key points at a row that exists, or the constraint has been violated. Neither says anything about anyone's character. Both describe whether a system's internal references hold.

Structural integrity   every member carries the load it claims
                       → test it, and the answer is yes or no

Referential integrity  every key points at a row that exists
                       → the constraint holds, or it is violated

REFerence [IN]tegrity  every citation points at a source that
                       exists and is the one it names
                       → the ledger balances, or it does not

A bibliography is a reference system in exactly the database sense. Each in-text citation is a pointer; each entry is the row it should resolve to. When a pointer resolves to nothing, that is not a moral failure, it is a broken reference — the same category of fault as a dangling key, and fixed the same way.

Which is why the word does the work no alternative does. "Accuracy" suggests the sources are correct, which is a claim about the literature. "Completeness" suggests nothing is missing, which is only one of the three claims below. Integrity is the word for a system whose internal references hold, and that is precisely what a bibliography either has or lacks.

Three claims, or it means nothing

A property that cannot be checked is marketing. So REFerence [IN]tegrity is defined as three claims that can each be shown to be true or false:

One — nothing was lost. Every citation read out of the document is accounted for in the result. Not "most", not "the ones we understood": every one. We hold this as a ledger, and it has to balance.

read from the document          412
                                ───
  matched to a reference        268
  in the bibliography only      121
  set aside as not-a-citation    23
                                ───
accounted for                   412      ✓  lost: 0

This sounds trivial and is not. The failure mode it prevents is the quiet one: a parser that cannot read a citation and simply drops it, producing a report that looks clean because the difficult items were never in it. If the ledger does not close, we have a bug — and we would rather find it than ship a tidy report.

Two — every match is evidenced. A pairing is not asserted because two strings looked alike. It is resolved against a real record and shown to you with what it was resolved against, so you can disagree. The Okafor case above is exactly what this claim exists to catch: the surname and year agreed, and the evidence did not. One author, two spellings is the same claim under the opposite load: the surname disagrees and the source is identical.

Three — you decide. Nothing is written into your document without your click. Not one correction, however obvious it looks to us.

The mark, and why the brackets

A trademark is a compression. It has to carry the whole idea in a glance, and the bracket is doing that work rather than decorating it.

Written in full the property is REFerence [IN]tegrity, and the capitals are not styling. REF and IN are the two halves of what is being claimed: a reference, and the integrity that has to be inside it rather than asserted about it from outside. Compressed, that is Ref[In] — the same two halves, with the brackets kept.

The brackets survive the compression on purpose. They are the punctuation of citation itself: [12] in Vancouver, the parenthetical in author–date, the square brackets an editor uses to mark an intervention in quoted text. A reader in this field sees a bracketed token and already knows it is a pointer to a source. The mark borrows that recognition rather than explaining itself.

And there is a practical reason to keep the setting consistent everywhere the words appear, which we do rigidly, including in this sentence. A brand that is set one way in a report, another in a heading and a third in a footer is not a mark — it is a phrase somebody keeps retyping. Integrity in the small things is the cheapest possible demonstration of the argument.

The family of terms, and what each one is for

Once the property has a name, the pieces that measure it need names too. They are not decorative — each one answers a different question.

The Ref[In] Report is the artefact: a document you can hand to a supervisor showing what was checked, what matched, what did not, and what was set aside. It is the thing that makes the claim auditable by someone who did not run it.

The Ref[In] Score summarises a single document on a fixed scale, so the same thesis checked twice is comparable to itself. It is a reading, not a grade — which matters, and we come back to it below.

The Ref[In] Architecture and Ref[In] Methodology are how the checking is done and why it is done that way — deterministic engines first, settling everything that can be settled by rule, with AI used only on what is genuinely ambiguous and never as the final word. Ref[In] AI is that constrained role, and its constraint is the point: it adjudicates, it does not invent. It has no mechanism for producing a reference that does not exist in a record it can cite back to you.

To be explicit, because it is the question we are asked most: your document is not training data. We do not train models on the theses that pass through the tool. The AI we use is a reasoning step over extracts, not a learner over your work.

Three levels, and what each one certifies

A document does not arrive at REFerence [IN]tegrity in one step, and the product does not pretend it does. There are three milestones, each meaning something specific, and the level only ever goes up.

Level 1   ORGANISING
          the cross-check has run: every citation harvested,
          every entry parsed, the two compared
          → you know what is missing and what is orphaned

Level 2   FINALISE
          the reference phase was reached: the gaps have been
          worked, sources resolved against real records
          → the list is now defensible, entry by entry

Level 3   EXPORT
          the export files exist: annotated paper, annotated
          bibliography, every reference carrying a verified
          link, the whole thing sealed and delivered
          → the claim is not just true, it is evidenced

Level 1 is diagnosis and it is where most of the surprise lives — it is the point at which someone learns their bibliography has nine citations pointing at nothing. Level 2 is the work. Level 3 is the artefact: the thing you can hand to a supervisor, an editor or an examiner, in which every claim the earlier levels made is shown with its evidence attached.

The reason to name them rather than leave them implicit is that they answer different questions and people conflate them constantly. "Have you checked your references?" is a Level 1 question. "Are they sound?" is Level 2. "Can you show me?" is Level 3, and it is the only one an institution can act on — which is why the annotation stage that produces it will not begin until the ledger from Level 1 balances. A certificate over an unbalanced ledger would certify nothing.

None of the three is a claim about your scholarship, and it is worth repeating in this context because the word "certificate" invites the misreading. A Level 3 document says your references resolve and here is the proof. It says nothing about whether they were the right sources or whether your argument holds — that judgement stays where it belongs.

What the name deliberately does not mean

Integrity is a loaded word, and we chose it knowing that. So the boundary needs stating plainly.

It measures the document, and it does not judge you. Most of what a check surfaces is ordinary: a citation added during a late edit and never carried into the bibliography, a surname transliterated two ways across four years, an edition superseded while the chapter sat with a supervisor. These are the marks of a long project, and a score that treated them as anything else would be measuring the wrong thing.

There is one finding that is not in that category, and we will not pretend otherwise. A citation to a source that does not exist — offered in support of a claim — is academic misconduct, and it is the one thing on this list that a reference check can actually establish. What follows from it is a matter for you and your institution. Our part ends at showing you, with evidence, that no record answers to it.

We built for that reality deliberately. Everything surfaces as a proposal you accept or reject, the report is written to be shown rather than to accuse, and there is no view anywhere in the product that ranks people. The unit of analysis is a document.

Referencing errors are what a four-year project does to a bibliography. The useful question is never "who is at fault" — it is "what does this list need before it is submitted".

Nor is it a claim about your scholarship. A sound reference list says your sources are correctly identified and correctly attached to your claims. It says nothing about whether they were the right sources, or whether your argument holds. That judgement belongs to you and your examiners, and no score should be allowed to imply otherwise.

What it looks like when the claim fails

A property worth naming is one that can be shown to be false, so it is worth being specific about what failure looks like for each of the three claims. None of these is hypothetical; all three are shapes we see.

Claim 1   nothing was lost
broken by  a parser that cannot read a citation and drops
           it — the report looks clean because the hard
           items were never in it
detected   the ledger does not balance: read 412,
           accounted for 389

Claim 2   every match is evidenced
broken by  a pairing asserted on surname and year with no
           record behind it
detected   the pair carries no resolved record, so it is
           shown as unverified rather than matched

Claim 3   you decide
broken by  a correction applied because the system was
           confident
detected   it cannot be: the change is a proposal until a
           person accepts it, and the prior value is kept

The third row is the one people find surprising, because every other tool in a writer's life applies its own suggestions and asks forgiveness. Here the refusal is structural: there is no path that writes into your document without a click, and the original value is retained beside every proposal so a wrong acceptance is reversible.

Why a name matters at all

Because the alternative is what we had: everyone describing the same property with different words, none of them meaning anything specific enough to test. "Clean bibliography" is a feeling. "The check passed" is a statement about software. REFerence [IN]tegrity is three claims and a ledger that closes — a thing you can hold us to, and a thing that can be shown to be false.

That is the whole ambition of the mark. Not to own a phrase, but to make one precise enough to be worth arguing about.