The seven stages of a citation check

A cross-check is not one operation. It is seven, in a fixed order, and each one is allowed to see only what the stage before it settled. Knowing the sequence tells you what any given screen is asking of you.

Most of the confusion people have with reference checkers comes from not knowing which question is currently on the table. This is the whole pipeline, in order, with what each stage decides and what it hands on.

  1  EXTRACT      read the document, collect every
                  in-text citation and every entry

  2  SCREEN       separate citations from things that
                  merely resemble them

  3  REVIEW       judge how well each reference entry
                  was parsed, correct from raw text only

  4  DE-DUPLICATE one work cited many ways becomes one
                  work, with its occurrences kept

  5  CROSS-CHECK  pair citations to entries; whatever
                  the rules cannot settle goes to review

  6  VERDICTS     matched · missing · orphan · unresolved
                  and the two paths out of them

  7  REPORT       what was measured, how, and the
                  Ref[In] Score

Why the order is fixed

Each stage depends on the one before it having finished, and finished means settled rather than merely run.

De-duplication cannot happen before parsing, because you cannot tell that two entries are the same work until you have read both properly. Cross-checking cannot happen before de-duplication, or one work cited three ways produces three separate verdicts about the same thing. The report cannot be written until the verdicts exist, because a report is a statement about verdicts.

Run them out of order and nothing errors. You simply get answers to questions that were not ready to be asked, which is the harder failure to notice.

Stage 1 and 2 are one decision, split

Extraction is deliberately generous: anything resembling a citation is collected. Screening then separates. Splitting them is what makes the generosity safe — a row set aside in stage 2 is still there, visible and reversible, where a row never collected in stage 1 would not exist at all. That asymmetry is the subject of the checker said 90% of my citations failed.

Stage 3 corrects the reading, never the source

The reference review looks at each bibliography entry and asks how well the parser read it — did it find the title, the year, the journal, the volume. Corrections are permitted only where they are recoverable from the raw text of the entry itself. If the entry has no title, the title stays missing; nothing is supplied from outside.

This is the stage people expect to be the AI's main job, and it is the most constrained one in the whole pipeline.

Stage 5 is where most of the value is

The cross-check pairs in-text citations with reference entries: shared identifiers first, then exact author and year, then the reference line as written. What the rules settle is settled. What survives goes to the AI match review as an explicitly marked residue.

The output is not a score at this point. It is four piles.

Stage 6: four verdicts, and what each one means

  MATCHED     cited in the text, listed in the
              bibliography, paired
              -> nothing to do

  MISSING     cited in the text, no entry found
              -> add the entry, or find it

  ORPHAN      listed in the bibliography, never
              cited anywhere in the text
              -> cite it, or remove it

  UNRESOLVED  a pair the rules could not settle
              and you have not yet judged
              -> read it, decide

Two of those four are honest and normal. A healthy thesis usually has a small number of orphans — a work read, listed, and in the end not cited — and a small number of missing entries from a late edit. A report of zeros is rarer than people expect, and is not the target. What a good report looks like is covered in one orphan, one missing.

Stage 4 is the one nobody expects to matter

De-duplication sounds like housekeeping and is closer to arithmetic. One work can appear in a bibliography several times — added once from a reference manager, once by hand, once more after a supervisor sent the same paper again — and the three entries will not be identical, or they would have been spotted already.

Left alone, that single work produces three verdicts. Perhaps one matched and two orphans, because the citations in your text all paired with whichever entry the rules reached first. Your report then shows two orphan references that are not orphans and a bibliography that looks careless when it is merely repetitive.

Collapsing them first means the work is judged once. The occurrences are kept — you can still see every place it was cited — but there is one entry, one verdict, and one thing to fix if it needs fixing. Getting this wrong does not produce an error; it produces a report that is quietly, plausibly wrong about how many problems you have.

Stage 7 reports only what it measured

The final stage assembles the document: counts, the four piles, how each pair was matched, which rows were reinstated during screening and by whom, and the Ref[In] Score.

The discipline in this stage is subtractive. It does not print the raw screened pile, it does not print prose about work it did not do, and where a figure was not measured it says so instead of estimating. A certificate that claims more than it checked is worse than a shorter one.

The seven stages above explain the work, rather than seven identical buttons or a promise that every optional service is free. Uploading, extraction and deterministic cross-check are free; optional AI review and other paid services follow the current pricing shown before you use them. You can review the basic findings before choosing whether to continue into paid work.

Where the two branches attach

Finding missing references and updating outdated ones are not stages in this list. They are separate runs that begin from a finished cross-check, because both need the verdicts before they have anything to act on — Find works from the missing pile, Update works from the matched pile.

That is why the tool asks you to complete a cross-check first. It is not a funnel. It is that neither branch has a question to answer until stage 6 has produced one.

What this does not claim

Seven stages is a description of this pipeline, not a standard. Other tools divide the work differently and some skip stages entirely — a checker with no screening stage is doing stage 1 and stage 2 at once and hiding the choice.

And the output of all seven is a statement about referencing quality: whether citations and entries line up with each other. It is not a similarity check, not an assessment of your argument, and not a finding of misconduct. A missing entry is nearly always a chapter reordered late, not a person concealing something.