What we make: all our sites and tools · How the lab ships code
Arc Codex
News stories with AI-written analysis beside them: the facts, a summary, and a longer take you can argue with.
Why it exists
What Arc Codex is for, who it is for, and how one story gets analyzed.
Read moreHow to use it
How to read the feed, the three kinds of analysis and the 48 A.R.C. patterns, and how to translate and share a story.
Read moreHow it is built
The software, the machines and the design choices.
Read moreContact
Report an error, ask a question, or ask about consulting work.
Read moreWhat we publish about ourselves
How stories get picked, what the scores mean, and the signed public record of what we published.
Read moreSputnik Watch
Sputnik Watch is an optional mode of Arc Codex's automated pipeline, switched on by the site's editor. It is paused as of 31 October 2026. When on, it answers information operations aimed at America and Europe: when Sputnik, a Russian state media outlet, publishes a story about the United States or Europe, our system writes and publishes a rebuttal, usually within minutes.
Every rebuttal is machine-written and labelled as such. It links the Sputnik original, quotes it only briefly, and may use only facts from Arc's archive of other outlets (never state media) or facts no reputable outlet disputes. Where the evidence is silent, it says so and explains the framing instead of inventing an answer. It states Sputnik's claim fairly and concedes what is accurate.
Before anything is published, a separate model grades the rebuttal against this checklist, and the score is printed at the end of every Sputnik Watch article:
- States Sputnik's claim fairly, in its strongest honest form (20 points)
- Every factual statement is cited to the evidence or is undisputed common knowledge (25)
- No invented facts, numbers, quotes, people or sources (25)
- Concedes what is accurate; separates what is reported from how it is framed (15)
- Calm and respectful: no mockery, insults or sweeping claims about Russians (15)
A rebuttal is published only with 70 points or more and no invented facts; the rest are held, not published. Stories older than a day are not answered. The grader is itself a small model and can be wrong: if you find an error, tell us through the contact page.
What's new
About pages corrected
We checked these pages against how the site actually runs and fixed what was out of date. Sentinel is off for new stories, narration recordings no longer speak a credit, feed stories may be translated by a cloud model, and our code is not public yet. Earlier versions said otherwise.
Short videos of some stories
Some stories now have a short captioned video, read by an AI voice from an AI-written script. Each is checked against the source by a second AI before it is posted, and that check can miss errors.
HOT channel
HOT writes stories on what Americans are searching for. An AI writes them from news reports, and a second AI checks them against those reports. Sputnik Watch is paused.
A signed public log
Each public article is now signed and added to a public log, so anyone can check that an article has not been changed without a record. See the Open Ledger page.
Sponsored-content signals compare writing style
For English-language stories, the sponsored-content section now compares the writing style with earlier stories from the same publication, once at least twenty are on file.
Sponsored-content signals now include links
For newly collected stories, the sponsored-content section also shows where the links in the story lead: public and research sources, one company's pages, tracking codes, or sales pages.
Sponsored-content signals under each story
New stories carry a section that counts three patterns common in paid placements and vendor press releases: who is quoted, subscription-style wording, and outside statistics followed by a product. They are shown with the sentences behind them, as evidence rather than a verdict. The About scoring page explains how they are counted.
Newsradio categories run longer
The newsradio station now keeps the last seven days of stories, so each category plays for longer. Programmes still start with the newest story.
Writing modes on the publish page
Write Prompt now asks for a purpose: Rebut, Promote, Explain or Free. Promotional pieces are labelled as such and are shared by direct link only.
Newsradio now plays only the latest day of news
The newsradio station now keeps just the last twenty-four hours of stories, so everything it plays is current news.
Newsradio now keeps about eighteen days of stories
The newsradio station now keeps the last eighteen days or so of stories to play, up from ten, so there are roughly twice as many to listen to.
Newsradio keeps ten days of stories
The newsradio station now keeps the last ten days of stories to play, up from two, so there is more to listen to between new stories.
The Quant speaks only when it has a point
Our numbers analyst now comments only when it can name a specific figure in the story and say what is missing from how it is presented, such as the year a change is measured from. Otherwise it stays silent.
Objectivity dial on non-English stories
Stories in other languages that have no objectivity score now show an "English only" note where the dial would be, instead of leaving a gap.
Story cards name the publisher
Each story card shows "Reporting by" the publisher under the headline and links to the original. Translated views say which language they came from, and images carry a caption. How we treat publishers' work is described on the Sources & credit page.