Citations that are checked before they reach you

Upload and cite, step by step

1. Sign in to Scholaris, or try it without an account in the demo.
2. In the Library (Biblioteca), drop the file (or paste an address: a web page, a PDF, YouTube, a podcast). The first pages of a digital PDF are searchable in under a second; those of a scan, in a few seconds.
3. In Search (Buscar), type the idea in your own words or the exact phrase in quotes.
4. In the result, copy the reference: it comes out in your style, with the printed page or the minute, as formatted text (italics included).
5. For a whole text, go to Write (Escribir) and paste your draft or upload your .docx: autocite proposes a citation for each claim, with its page, and you accept or discard them one by one.
6. Export the cited text (.docx, Markdown, plain text or LaTeX) and the bibliography (BibTeX, RIS or CSL-JSON).

Styles

Citations are formatted by citeproc-js with the official CSL files, in four locales (es-ES, en-US, fr-FR, it-IT). The default style is APA.

When the date of the original work differs from the edition's, it appears as the original date ("1605/2004"); the edition is given unless it is a same-year reprint.

Where each citation comes from

No model writes the citation. The locator ("p. 23", "12:04", "slide 7") comes from the anchor stored when the document was read (see Formats and anchors), and author, year and title come from the record, whose provenance you can inspect field by field. The record is checked against Crossref, OpenAlex, Open Library and Wikidata, and an external record is only accepted if the title matches and the author or the year matches too.

Autocite

For a whole draft, Scholaris:

1. splits it into paragraphs and claims;
2. retrieves fourteen candidate passages for each;
3. lets a model propose which passage supports each claim;
4. checks in code that the evidence appears verbatim in the passage, that a negative statement is not cited as if it were positive, that the key terms are there and that the chronology is possible;
5. and runs every proposal past a judge (Jev, by TypeSafe) in batches of eight.

Proposals with probability 0.7 or more are accepted; between 0.4 and 0.7, flagged for review; below that, discarded. Consecutive verified pages are merged ("pp. 23-25"). In the benchmark, precision was 95.1 %, recall 96.7 % and invented citations zero.

Verifying a claim

You give it a sentence and Scholaris says whether your library supports it, with the passages:

respaldada (supported)

the best passage supports it directly, with probability 0.7 or more.

parcial (partial)

some passage partly supports it (probability 0.4 or more).

contradicha (contradicted)

some passage says the opposite, with probability 0.5 or more.

sin_respaldo (unsupported)

no passage supports it.

The judge also classifies how each passage relates to the claim: direct support, application of a framework, context, contradiction, reported opinion or negative statement.

Temporal logic

Comparing dates is not left to a model: code does it.

• The year of the original work is used, not the edition's.
• A text cannot directly support a claim about something invented later: Scholaris knows when terms like "internet" (1983), "deconstruction" (1967) or "transformer" (2017) appeared.
• If the claim says one author "anticipated" or "built on" another, the chronology has to be possible; if it is not, the citation is dropped.
• If the claim refers to a year more than twenty years after the source, direct support is downgraded to "application of a framework".
• A source later than the year of your own text cannot be cited in it.
• Undated sources ("n.d.") skip all of this.

Inside Word

Scholaris opens the .docx, inserts each citation in place keeping the formatting of every run of text, uses real Word footnotes for note styles and appends the bibliography. It also takes ODT, plain text, Markdown and HTML, up to 50 MB.

Copying a reference

Every document has a button to copy its reference in your style (or another, with a preview). It is copied as formatted text, italics included, and as plain text.

Export and import

• Export: BibTeX, RIS and CSL-JSON; the bibliography as text, HTML or Markdown.
• Import a BibTeX file (from Zotero, say): each entry is matched to your documents by DOI, then ISBN, then title; those that do not match stay as references without text.