What happens before anything is published

Quill’s articles are drafted by a language model. Models invent things, so the question is not whether the drafting is trustworthy — it is not — but what has to be true of a draft before it can become an article. Five things, in order.

1. The draft can only see stored sources

Before drafting begins, an editor fixes the set of sources the article will be grounded on: full texts already harvested into Quill’s own store, not search results fetched while writing. The model is given those texts. It has nothing else to work from, and the set is recorded with the article — it is the list you see in the evidence rail beside the prose.

2. Every factual claim arrives with a quote

The draft is not prose alone. Each factual claim is emitted alongside the source it came from and the verbatim passage that supports it. A claim with no such record does not become a claim; it becomes a flag.

3. The quote is matched against the stored text

This check involves no model at all. The cited source must be in the grounding set, and the supporting quote must match the stored text of that source. A fabricated quote fails a string comparison, which is the point: a plausible invention and a real passage are indistinguishable to a reader and trivially distinguishable to a matcher. The match forgives the wobble of copying (spacing, punctuation, a slipped word), never a changed figure or an added or dropped negation.

4. A second model argues the other way

Matching a quote proves the quote exists, not that it supports what the claim says about it. So a separate pass reads claim and passage together and asks whether the passage actually bears the weight — strictest where the claim carries a figure, which is where drift is hardest to notice and most damaging.

5. An editor clears it, or it does not go out

Everything above surfaces to a human as flags against specific sentences, beside the source passage in question. A mechanical failure blocks publication. The checks are run again at the moment of publishing, against the sources as they stand then, and they cover the headline and standfirst as well as the body. Judgement calls, and any claim the second model could not judge, can be dismissed by the editor, with a reason recorded against the article.

Mechanical checks match strings and figures, and they are sometimes wrong: a unit written in capitals, or a figure split across a table, can make a true passage unfindable. When that happens an editor may override the failure, but only after confirming the claim against the source themselves. The override is recorded with their name, the time and their reason, and the claim is marked on the article: confirmed by an editor; the automatic check could not confirm it. So a mechanical failure never publishes silently. Either it was fixed, or a named person stood behind the claim and you can see which claims needed one. Nothing is published by the pipeline; publication is a decision a named person makes.

Sources are graded, and grading is not verification

Each source carries a tier: A peer-reviewed, B editorially reviewed, C vendor or ad-hoc material. Tier-A findings may be stated as findings. Tier-B and tier-C claims must be attributed in the prose as someone’s position rather than voiced as established fact, and vendor sources are marked attributed, not endorsed wherever they appear. In a research synthesis, a claim carrying a measurement must cite a peer-reviewed source. That is checked mechanically, against each source’s tier at the moment of publishing, and like any mechanical failure it blocks the article unless an editor overrides it, with the claim then marked as confirmed by an editor. A date — the year a paper appeared, the month an incident was disclosed — is held to a lower bar: an editor may publish it on an editorially-reviewed source, having said so.

What this does not establish

A verified claim is one that matches what its source says. It is not a claim Quill has established to be true. A peer-reviewed paper can be wrong; a correctly quoted vendor can be selling something; a set of sources chosen by an editor is a set of sources chosen by an editor. What Quill guarantees is that you can check the article against the record it was built from — every claim carries a numbered anchor to the source it came from, and the full source list travels with the article, including in the feed.

If a published article turns out to be wrong, it is corrected in place, re-verified, and dated. See About.