Extract the sources
The starting point is the finished or nearly finished report, as a PDF. Checked Sources reads the text of every page, including footnotes, and the link annotations behind clickable text, which often differ from the words shown on the page. Scanned pages are read with OCR (optical character recognition, which converts an image of text into text).
From that text it builds a list of items: quotations, legal citations, references to journal articles and books, DOIs (the permanent identifiers assigned to journal articles), and web addresses. Each item carries its page and footnote location and a confidence score for how it was linked to its source.
For confidential work there will be a no-upload mode. The report is read inside your own browser, and only the extracted list and a fingerprint of the document reach the service.
Correct the list
Automated extraction is never perfect. Before review begins, you can edit, merge, split, add or remove items. Duplicate URLs are flagged, while each distinct quotation keeps its own entry. The review does not start until you confirm the list.
Review each source
The review window presents one item at a time: the quotation or citation as it appears in the report, its location, and the source. Before you open a source, you name what you will check the item against, whether that is the cited page, a court opinion, a dataset or a printed book.
For web sources, Checked Sources retrieves the page itself, records the redirect chain and the server's response, and checks whether the quoted words appear on the page. Where the browser supports it, the source opens scrolled to the quoted passage. You then record your own finding: verified, discrepancy found, or not checkable, with a note.
Record every step
Each action becomes an event in an append-only log, meaning events can be added but never changed or removed. Each event carries a UTC timestamp and a hash, a short fingerprint computed from the event and the one before it. Altering any earlier entry would break every fingerprint after it.
Export the record
The export is a ZIP package containing a PDF report, a CSV log with one row per event, the captured pages, and a manifest of SHA-256 fingerprints for every file. A small checking script is included so that anyone, including an opposing expert, can confirm that nothing in the record changed after it was exported.
Every export also states its limits plainly. The record shows what was checked, against what, when and by whom. It does not prove that a reviewer understood a source, that a source is accurate, or that a report is admissible. Those questions belong to the reader, the client or the tribunal.
What is planned after the first release
- A browser extension that captures a page exactly as the reviewer sees it, including sources behind a login or paywall
- Independent timestamps from an RFC 3161 timestamp authority, an outside service that certifies a value existed at a given time
- Re-verification when a report is revised, carrying forward only the checks that still apply
- Team accounts that separate who extracted, who verified and who signed