A citation here is a document,
a line span, and the words themselves
Not a footnote, and not a link to a website. Every answer names the converted document it came from, gives the lines it was taken from, and repeats those lines word for word. The file is on your disk. You can open it and see the same words in the same place. If you cannot, the citation is wrong, and this page says when that happens.
Open the file. The words are there.
Most software that cites does so after the fact: it writes the answer, then looks for something to attach to it. Here the order is reversed. The citation exists first — a passage found in your documents — and the answer is that passage, cut to three lines. There is nothing to attach because nothing was written.
The three parts
id · locator · text| Part | What it is | What you do with it |
|---|---|---|
| Document id | The identity of the converted document the passage was taken from. One converted document per source file, and the document carries the path of the original it was read from. | Find the file. |
| Line locator | The line numbers in that document the passage occupies, written as a span —
lines a–b. |
Go to the place. |
| Verbatim text | The lines themselves, exactly as they appear in the document. At most three. | Read the words and decide whether they answer your question. |
A citation with any one of the three missing would not be a citation. An id without lines is a hint. Lines without the text are homework. Text without either is what everybody else ships.
You can open the file
and you shouldThe document the citation names is a readable Markdown copy sitting on your own disk, made by the conversion half of the product from a file you already had. The original is never altered; conversion writes a new file and leaves the source byte for byte as it was. So what you open is the copy, and the copy carries a note of where it came from and how it was read.
That last part matters when the source was paper. If the original was a scan, the words in the copy are what the machine read off the page, and machines misread paper. The archive OCR run done so far — 7,996 documents, 35,633 pages (counted) — was done entirely on the machine with nothing sent out. But a citation into a page read by OCR is a citation into the machine's reading, not the ink. When the figure matters, check it against the scan. And in the proving corpora, scanned paper is mostly not searchable yet: one company's OCR queue stands at 0 of 883 candidates read, and photographed documents are held as media only.
Three lines, two citations
a promise, not a limitThe engine will quote at most three lines and attach at most two citations. Those are constants in its source. They read like limits. They are a promise.
Three lines is what a person can check in the time it takes to open a file. A ten-line answer is a summary, and a summary is something the software composed — which is exactly what this product refuses to do. If your question needs more than three lines to answer, the right result is two citations to two places and you reading both, or a refusal. Not a longer paragraph.
Two citations is the same rule from the other side. An answer that needs five sources to hold it up is an argument, and the engine does not make arguments. It finds the best passage, and at most one more, and shows them to you.
The answer is a line out of your document, not a sentence about it. Every choice on this page comes back to that sentence.
The known bug
stated, not buriedA stitched answer is reported as one span
When the three lines the engine quotes are not next to each other in the document — say lines
12, 14 and 19 — the citation reports them as a single contiguous range, lines 12–19,
as though the whole stretch were the answer. It is not. The lines in between are not part of
what was quoted. The words are still verbatim and the document is still right, but the locator
overstates its span.
This is a known faithfulness bug. It is recorded in the engineering dossier as finding Q3 and it is not fixed as of this writing. What it means for you: when you open the file at a cited range and find lines that were not in the answer, that is this bug, not a different document. Read the quoted words against the range, and trust the words over the numbers until the locator is corrected. What is built, what is specified and what has never been run is on the record.
One more thing a citation does not tell you: who wrote the document, or whether it was right when they wrote it. A citation says where the words are. It does not say the words are true. Provenance — the path, the extraction method, the status, and for a scan the reading confidence — is the next page, and the difference between provenance and authorship is the point of it.
Read next
three doorsProvenance
Source path, extraction method, status. Why provenance is not authorship, and why that matters when you rely on a line.
Provenance ▸ The mechanismHow it answers
From question to quoted line: the three retrieval legs, the fusion, and the extraction that produces the citation.
How it answers ▸ The ledgerThe record
Every known shortfall, including this one, with what would close it. The page that makes the rest believable.
Read the record ▸