What the pipeline does, and the places it deliberately stops short.
LaTeX is the best input by a wide margin, because a \cite names
its bibliography entry exactly and its position is known to the character. LyX
and PDF both work; PDF has to infer what LaTeX simply states — columns are found
by locating the gutter, reference strings are split and parsed, and citations
are matched to entries by author surname and year. It loses citations it cannot
match confidently, and the report says how many rather than quietly proceeding.
Each reference is resolved to a canonical work: by DOI where there is one, then by identifier, then by title-and-author search across OpenAlex and Crossref. Every resolution carries a method and a confidence, and anything below the bar is reported as unresolved rather than guessed at.
Economics needs one thing most fields don't: a working paper and the journal article it became are the same work in two versions, and citation counts, publication dates, and retraction notices attach differently to each. Relit keeps both and records the relationship between them, so a 2018 article and the 2014 NBER paper behind it are neither conflated nor treated as strangers.
Around each citation, relit takes the surrounding sentences as the claim being made, with the citation itself marked so the model can tell it apart from others nearby. It then reads the cited work and asks whether that claim survives contact with the source.
Cheap models handle the easy cases and are only allowed to clear a citation, never to flag one; anything that looks like a problem escalates to a stronger model, and every flag is re-examined before it reaches you. An error in the cheap pass therefore costs a missed finding, not a false accusation.
Deciding that no prior work exists is a claim about the entire literature, and current models are bad at it while sounding certain. Relit surfaces candidate prior work and shows what it searched. Whether that work damages your contribution is a judgement it leaves to you.
Not finding a problem is not the same as establishing there isn't one, and the difference usually comes down to whether the source could be read at all. Every report carries what was reachable and what wasn't.
Text is fetched to locate quotations. Reports carry short quotes with locators. Paywalled material is skipped rather than worked around, and sources that permit only metadata are used only for metadata.