CONTEXT(CLONE)
Sign in

Check that every citation supports its sentence

For: Production editor or research integrity officer at a publisher or university press

Pattern: Map, verify, reduceNeeds scaleDesigned for 40 to 1,200 agents
Get early accessSign in to use

The pain today

A monograph or report carries hundreds of citations. Nobody checks that each cited source says what the sentence claims, and a misquoted source surfaces after publication.

The ask

“I attached the manuscript and the full texts of the cited sources we hold. For every citation, tell me whether the cited source supports the sentence, partly supports it, contradicts it or could not be found in the text, with the passage from the source.”

Plain words, as you would say it to a colleague. Edit it to fit your case before you send it.

What you attach or connect

  • Manuscript with reference list
  • Full texts of cited sources you hold
  • House style for citations

The unit of work

One worker task per one citation and its cited source.

Why a swarm fits

Each citation is one sentence and one source. Hundreds of them are fully independent, and no worker needs the rest of the book to check its own.

Not for

Sources you cannot supply as text. It does not fetch papers or judge whether a source itself is reliable.

The decision tree

6 typed decisions, each with an action for every answer

At fixed moments in a run, the engine puts one narrow question to a decision model. The decision model never writes text: it answers yes or no with a probability, picks from listed options, or gives a score, about a small slice of the material. The engine then does exactly what this tree says, which is what makes the run auditable. The thresholds are the template's design values, not measured results.

  1. Planner, while planning

    Scope checkA choice among options

    Before work starts on a unit

    Is the full text of the work named in this reference among the supplied sources, in the edition the reference cites?

    Sees only: One reference list entry and the title pages of candidate supplied sources

    Why: Citations that cannot be checked are listed as unchecked instead of being guessed at.

    • Supplied, same editionthenAccept
    • Supplied, different editionthenMark unresolved
    • Not suppliedthenSkip this unit

  2. Before workers, before a task runs

    Small worker or strong modelYes or no, with a probability

    Before a task runs

    Does the citing sentence attribute one plain statement to the source, rather than a synthesis, a comparison or a statistical result?

    Sees only: One citing sentence

    Why: Plain attributions are cheap to check; syntheses and statistics need the stronger reader.

    • Yes: 0.60 or higherthenAccept
    • Unsure: 0.40 up to 0.60thenEscalate to a strong model
    • No: below 0.40thenEscalate to a strong model

  3. After workers, the judge checks

    Evidence checkA choice among options

    After a worker answers

    Does the quoted passage in the cited source make the statement the manuscript attributes to it, rather than a weaker, narrower or different one?

    Sees only: One citing sentence and the passage found in the cited source

    Why: Overstated citations are the common failure, so partial support gets its own outcome.

    • Supports as statedthenAccept
    • Supports only a weaker claimthenMark unresolved
    • Says something different or oppositethenEscalate to a strong model
    • Passage is not relevantthenReject and retry

  4. Evidence checkYes or no, with a probability

    After a worker answers

    Do the words inside the manuscript's quotation marks match the source passage word for word, apart from marked omissions?

    Sees only: One direct quotation from the manuscript and the source passage

    Why: A misquotation in print is an erratum; this check is cheap insurance.

    • Yes: 0.90 or higherthenAccept
    • Unsure: 0.50 up to 0.90thenReject and retry
    • No: below 0.50thenMark unresolved

  5. Run control, between rounds

    Retry or stopYes or no, with a probability

    After a rejection or low confidence

    Is the source long or structured enough that the statement could sit in a chapter or section the worker did not search?

    Sees only: The source's table of contents and the sections already searched

    Why: A not found verdict is only fair after the whole source has had a chance.

    • Yes: 0.60 or higherthenEscalate to a strong model
    • Unsure: 0.30 up to 0.60thenEscalate to a strong model
    • No: below 0.30thenMark unresolved

  6. Accountable person, before anything is settled

    Person decidesYes or no, with a probability

    Before anything is reported as settled

    Does this citation's verdict say the source contradicts the sentence, supports only a weaker claim, or was misquoted?

    Sees only: One citation's verdict with both passages

    Why: The author and production editor decide corrections; patterns go to the research integrity officer.

    Accountable: The author and production editor decide every correction; integrity concerns go to the research integrity officer.

    • Yes: 0.40 or higherthenAsk a person
    • Unsure: 0.15 up to 0.40thenAsk a person
    • No: below 0.15thenAccept

The fleet: who does what

Model tiers by role, not brands: you choose the models. Strong reasoning models plan and reconcile, small fast models do the wide work, and the judge is a decision model from a different family, so it does not share the workers' blind spots.

  1. Planner

    A strong reasoning model extracts every citing sentence and pairs it with the right source text.

    Decisions here:1. Scope check

  2. Workers

    Small fast workers from an open-weight family, one per citation, finding the passage in the source.

    Designed for 40 to 1,200 agents, one worker task per one citation and its cited source. Each worker receives only its own unit.

    Decisions here:2. Small worker or strong model

  3. Judge, from a different model family

    A decision model from a different family asks whether the found passage supports the citing sentence.

    Decisions here:3. Evidence check4. Evidence check

  4. Reconciler

    A strong reasoning model lists citations by verdict per chapter and keeps borderline cases open.

    Decisions here:5. Retry or stop

  5. Accountable person

    The author and production editor decide every correction; integrity concerns go to the research integrity officer.

    Decisions here:6. Person decides

Checked before anything is accepted

  • Supports and contradicts verdicts must quote the source passage verbatim
  • Not found verdicts are rechecked with a wider search of the source
  • Direct quotations in the manuscript are compared word for word with the source
  • Citations whose source text was not supplied are listed as unchecked

What comes back

  • Verdict per citation: supports, partly, contradicts or not found
  • Source passages behind each verdict
  • Misquotations, with both versions
  • Unchecked citations and the sources needed

What to measure

  • Share of flagged citations the author or editor corrects
  • Citation errors reported after publication
  • Editor hours per chapter
  • Cost per citation checked

Names of measures only. No result is claimed for this template.

Templates open in the workspace chat with the ask filled in. Nothing runs until you send it.

Get early accessSign in to use

More in Education and publishing

Map course modules to accreditation learning outcomes

For: Programme director or quality assurance lead preparing an accreditation review

Accreditors ask where each required outcome is taught and assessed.

Pattern: Hierarchical decompositionNeeds scale6 decisionsDesigned for 30 to 700 agents