Relit — Every citation, read the way a referee would.

The gap

Reference checkers ask whether a citation exists. Increasingly they have to: fabricated references are a documented and growing problem, and several tools now catch them.

Almost nothing asks the harder question — whether the cited work supports the sentence it is attached to. That failure is quieter and probably far more common. The reference is real, the authors are real, the paper exists; it just does not say what the citing sentence claims it says. A reader who follows the citation finds out. A reader who doesn't, doesn't.

Relit is built for that second question.

What it does

Citation support

For each citation, retrieve the cited work, locate the relevant passage, and judge whether it supports the claim — supported, partially supported, unsupported, contradicted, or not checkable — with the quote it relied on.

In development

Bibliography and versions

Retractions, corrections, and expressions of concern. Duplicate and malformed entries. Working papers cited where a published version now exists — tracked as a link between versions, never a silent substitution.

Runs today

Reception

How the literature has actually taken up each work you rely on: citation counts, journal rank, citations from leading journals, and rates measured from both first circulation and publication.

Runs today

Discovery

For explicit novelty and precedence claims, an adversarial search for counterexamples, reported as candidates with a record of what was searched. Never as a verdict.

In development

Evidence, or it doesn't ship

A tool that flags citation problems is only useful if its flags are right. Confidently wrong findings would waste an author's time and, repeated, would be worse than no tool at all. Four rules follow from that, and they are enforced in the code rather than left to good intentions:

Every finding carries its evidence. A manuscript location, and where the finding rests on a cited work, a verbatim quote with a locator you can check.
A verdict is capped by what we could read. If only the abstract was reachable, that can support a mild observation but never a confident contradiction — it becomes not checkable instead.
Identity is settled before anything is judged. A reference matched only by approximate title similarity is never verified against, because a misidentified work would produce a citation complaint about a paper you never cited.
Silence is reported, not implied. Coverage and what could not be read are part of the report. "No findings" must never be mistaken for "checked and clean".

How it works in more detail →

Where it stands

The deterministic half runs: ingestion from LaTeX, LyX, and PDF; bibliography parsing; reference resolution against OpenAlex and Crossref; working-paper to published-version linking; the hygiene checks; and the reception table. Citation support and discovery need models and are being built now.

Relit is not open for general use yet, and no results from it have been published. Numbers will follow the benchmark, not precede it.

If you write in economics or a nearby field and would want this run on a working paper, write to [email protected] — early access is going to a small number of authors first.