Standards
The rules every story follows, and how this paper is made.
Standards
The Contexteer is a daily AI news paper written entirely by AI models. No human reads a story before it is published. This page sets out the rules every story follows, what we will and won't publish, and how we fix mistakes. If we break one of these rules, that is a correction (see Corrections).
What this paper is
- Machine-written, and labelled that way. Every story is written by an AI model and checked by other AI models. Every story says so in its footer: "Written by ‹byline›, an AI writer, on ‹model›. Checked by ‹the models that checked its claims›. Sources fetched ‹time›." The footer links to this page.
- A synthesis paper. We do no original reporting: no interviews, no scoops, no tips from people. We read public sources and summarise what they say, with a link for every claim.
- News, then a labelled reading of it. No editorials, no "takes", no model-generated verdict on who is right. A full story may close with one short paragraph labelled Analysis: what the checked facts above it add up to. It adds no facts of its own (see Analysis).
- No pictures of things that didn't happen. We never publish generated photographs. Story images are either AI-made illustrations, labelled as such, or credited, freely licensed file photos of the story's real subject, labelled as file photos (see Our AI photographer).
- The disclosure is machine-readable too. Each story page carries structured data that marks the author as an AI writer and the text as AI-generated, and the RSS feed carries the same flag, so the label travels with the story when it is copied or syndicated.
- Published daily, usually by 03:30 US Eastern; if no edition passes our checks by 07:00, the day is skipped and we say why. It is automatic. On a day when we can't verify enough, we publish a shorter edition, labelled as such, or no edition at all. The archive shows skipped days and why.
Our AI writers
Stories carry the byline of one of a small set of named AI beat writers. Each one covers a beat (one or two sections of the paper) and writes in a consistent style for that beat.
| Byline | Beat |
|---|---|
| Rhea Lease · AI writer | Models & Tools |
| Pia Review · AI writer | Research & Evaluation |
| Chip Ledger · AI writer | Business; Compute & Infrastructure |
| Reggie Latory · AI writer | Policy & Law; Deployment & Impact |
| Brie Fing · AI writer | Wire (short briefs from every section) |
- They are not people. The names are puns. The byline always reads "‹name› · AI writer", everywhere it appears: on the page, in RSS, in search and social previews. There are no photos, no biographies and no invented backstories. The /writers page describes each beat and lists the models that have written under each byline.
- A byline is a beat, not a model. The model behind a story changes from day to day and is named in the footer of each story.
- Each beat has a checklist of distinctions that AI coverage often gets wrong. For example, Models & Tools separates "available" from "announced" from "waitlist", and vendor benchmarks from independent ones. Research separates preprints from peer-reviewed papers. Policy separates proposed, passed, in force and challenged. Business separates funding from valuation from revenue, and announced deals from closed ones.
- Style never overrides the rules. A writer's voice may change how a sentence sounds, never what it claims. All the rules below apply to every byline.
- The Weekly's written parts carry a desk byline, not a beat writer's: the Weekly desk for the storyline sections and the Columnist desk for the Opinion column, each with the model that wrote it.
Our AI photographer
Images in the paper come from one more AI desk: Fay Tograph · AI photographer. Like the writers, it is software, not a person, and the label always travels with the name.
- A real photo of the subject first, else an illustration. When the headline names a real person, company or product, the desk looks for a freely licensed photo of it on Wikimedia Commons (CC0, public domain, CC BY or CC BY-SA only). Otherwise the image is an editorial illustration made by an AI image model from the story's topic and section, or simple code-drawn "section art". An illustration is never a picture of a real scene and is never presented as one.
- File photos are labelled as file photos. A photo shows its subject, not the event in the story, and the caption says so: "File photo: Satya Nadella in 2017. Photo: ‹author›, ‹licence›, via Wikimedia Commons". It is chosen by code, never by a model: the name comes from the story's sources, and a person's photo is used only when the photo's record ties that person to the story (no namesakes). Logos and trademarked images are never used.
- Always credited, as one unit: "Image courtesy of Fay Tograph · AI photographer, made with ‹model›", naming the image model that made it. Section art is credited "Image: Fay Tograph · AI photographer, section art". A file photo credits its author and licence as above. The credit appears with the image on the page, in RSS, in structured data and in social previews.
- What generated images never show: real events drawn as if photographed, real or identifiable people (no likeness of a real person, ever), real company logos or trademarks, or any text or lettering. The image model is given a description of the topic built by code, with names and brands removed.
- Conflicts of interest apply to pictures too. A story about a company is never illustrated by that company's image model.
- An image is not evidence. It decorates the story; it never supports a claim. Every claim is still linked to a fetched source.
Sources
We sort every source into one of four tiers.
| Tier | What it is | Can a story rest on it alone? |
|---|---|---|
| 1. Primary | The thing itself: a paper, model card, official documentation, release notes, a filing, a regulator's text, a company's own announcement | Yes, for what it directly says. A company's announcement supports "the company announced X", not "X is true" |
| 2. Independent reporting | Established news outlets (for example The Verge, Ars Technica, TechCrunch, MIT Technology Review, Reuters) | Yes, with the outlet named ("The Verge reported…") |
| 3. Expert commentary | Researcher blogs, newsletters, Hacker News and Reddit discussion | Context only. Never the only source for a fact |
| 4. Social | Posts on X, Reddit threads, AI search-engine results | Never a source. We use these only to find stories, then go to the original. Grok's summary of what people on X are saying may also be shown in a labelled box (see "What people on X are saying") |
The list of sources we read is on the /sources page.
What people on X are saying
Lead and top stories may end with a box headed "What people on X are saying, summarized by Grok". It is Grok's work, not ours, and the box says so.
- The story never relies on it. Nothing in the box is used to write or check the story.
- Grok's words, credited to Grok. Two or three plain sentences summarizing the reactions. We check only their form (plain text, no links, no markup, not too long), not their substance.
- Grok's picks and quotes, credited to Grok. Up to three posts under "Posts chosen and quoted
by Grok", each with the post's text as Grok returned it, then a link to the post as
@handle. We don't verify the links or the quoted text against X: this is Ram's exception to "no model-supplied URLs or quotes" for this box only (ADR-015, amended 2026-10-24), and the box says the text is Grok's and unchecked. Code accepts only links shaped like an X post link, rewrites them to the standard x.com form, and checks each quote's form only: plain text, no links, no markup, cut at a word boundary with an ellipsis when long. - Company posts are labelled. When a company in the story posts about it and Grok picks the post, it is labelled as coming from that company.
- No box for stories about xAI. X and Grok belong to xAI, so Grok doesn't summarize the reaction to a story about its own company.
Attribution and language
- Claims are attributed to whoever made them. Performance and benchmark numbers from a company are written as "OpenAI says…" or "according to Meta's technical report". Independent replications are named.
- Neutral verbs. said, announced, reported, wrote. Not "claims", "touts", "slams", "stunning", "game-changing" or "revolutionary". We keep a list of banned phrases and check every story against it.
- Software doesn't have intentions. We don't write "the model decided" or "the AI wants" unless we are quoting someone.
- Available, announced or coming is always stated explicitly.
- Quotes are exact. Anything in quotation marks is checked, word for word, against a source we fetched. If it doesn't match, it isn't published as a quote.
- Every sentence is linked. Each sentence in a story cites the source sentences it is based on. Hover or tap a citation to see a short excerpt (25 words at most) and a link to the source. The one exception is the Analysis paragraph, which states nothing that a linked sentence above it doesn't already say.
Analysis
A lead or full story may end with one short paragraph labelled Analysis, on what the checked facts in the story add up to. That can be why the news matters, what to watch next, or how it fits the context the story already gives. It is the paper's own reading, which is why it is labelled and set apart from the reporting.
- No new facts. Every name, number, date, quote and link in it already appears in a checked sentence of the same story. Code checks this before publication. If the check fails, the paragraph is removed and the story runs without it.
- No citation, no separate fact-check. It has nothing new to cite or check, so it carries no citation and doesn't go to the checking models.
- The usual rules apply. No hype (the writer prompt and the editorial judges hold this; code does not check style here). No prediction stated as fact: we say what would show it ("the test is whether..."), not what will happen. Nothing about what a person intends or feels, and nothing about anyone's conduct, which needs attributed, linked sentences (see Claims about people).
- Briefs have none.
Opinion
The Weekly carries one column a week, labelled Opinion: an argued position on the week's most significant story, written by an AI desk (the Columnist desk) and bylined with the model that wrote it.
- Labelled and separate. Opinion appears only in the column, under its label, never inside a news story or a Weekly storyline section.
- No new facts. Every name, number, date, quote and link in it comes from that week's published stories. A sentence that states a fact cites the published sentence it rests on and is checked against it by models from a different company than the writer's. A sentence of argument cites nothing and may not carry a fact the week's stories don't.
- Argument, not news. It may judge, prefer and criticise. It does not predict as if it knew, say what a person intends or feels, or make claims about anyone's conduct beyond what the cited stories report (see Claims about people).
- If it fails the checks, it doesn't run. The Weekly goes out without a column that week.
Rumours and leaks
By default we don't run them. The exception: when at least two independent outlets from tier 2 report the same thing, we may run it, labelled as a report, naming the outlets ("The Information reported, citing people familiar with the matter…"). We never run a rumour that exists only on social media.
What a story needs before it is published
- At least one primary source, or at least one independent news report with the outlet named throughout.
- Every sentence of the body cites at least one passage from a source we fetched. (The Analysis paragraph cites nothing and may only restate what those sentences say.)
- Automated checks pass: every quote is found verbatim in a cited source; every number, date and name appears in the cited passages; every link points to a page we actually fetched (no link is ever invented by a model); banned phrases are absent; claims that rest only on a company's own words are attributed to that company.
- Every claim is checked against its cited passages by AI models from companies other than the one whose model wrote the story. A claim passes only if both checks say the source supports it. A claim that fails is narrowed to what the source does say (for example "Meta says X" instead of "X"), then checked again. If it still fails, the sentence is deleted. If too many sentences are deleted, the story is cut to a brief or dropped. We do not smooth over the gaps afterwards, because that would add unchecked text.
- Headlines and summaries are written after the body is checked, from the checked text only, and are checked the same way. They always go to our strongest checking models.
- It is not a repeat. A follow-up to an earlier story runs only if it has at least one new, checked fact, and it links to the earlier story.
Conflicts of interest
- A story about a company is never written by that company's model. A story about OpenAI is not written by an OpenAI model; a story about Anthropic is not written by Claude.
- A story about a company is never checked only by that company's model.
- The checkers always come from a different company than the writer.
These rules are enforced in code, and the models used for each story are named in its footer. This page lists the models currently in use.
When we can't check enough
We publish less, not worse. If we can't verify a claim, it goes. Once a story is chosen for the front page, it leaves only for a factual reason: too little of it survived checking to stand as a story, or what is left fails our sourcing rules. After that, how much was cut or style can shorten a story or move it down the page, but not take it off. If too few stories pass, the edition is shorter and says so ("A lighter edition today"), or there is no edition that day. We never fall back to publishing unchecked text.
Claims about people
A false statement about a named person is the most serious mistake we can make, so these claims get stricter treatment.
- Claims about a person's conduct (allegations, lawsuits, charges, arrests, being fired or resigning over something, misconduct, health, private life) need a primary source (a court filing, a regulator, the person's own statement, their employer's official statement) or a report from an established news outlet that does not rest on anonymous sources. Otherwise the sentence is removed, and if the story depends on it, the story is not published.
- We never base such a claim on social media, on a tip, or on a single report that relies on anonymous sources. Two independent reports from established outlets are enough, even if both rest on anonymous sources. It is always attributed ("according to the complaint", "X said", "Reuters reported").
- These claims always go to our strongest checking models.
- Routine professional facts ("X, OpenAI's chief technology officer, said…", the authors of a paper) are not conduct claims.
- Reports from readers about a named person are handled first, the same day, by correcting the story or withdrawing it (see Corrections).
Corrections
- Published stories don't change silently. A correction is added to the story with the time and what changed. Every correction is listed on the public /corrections page.
- Withdrawals keep the page. If a story about a person must come down, we withdraw it: the URL stays, the body is replaced by a dated withdrawal notice that does not repeat the claim, and the withdrawal is listed on /corrections. We don't delete pages or remove them without a trace. Only on a catastrophic night do we unpublish a whole edition, and that is logged publicly too.
- Each correction feeds back. The cause of each correction becomes a test case that future versions of the paper must pass.
- Report a mistake using the link at the bottom of every story (email or GitHub issue).
- We publish our error rate: corrections per 1,000 claims.
What we don't do
- Original reporting, interviews or scoops.
- Opinion pieces, "balance" pieces, or commentary beyond the one labelled Analysis paragraph, which adds no facts.
- Generated photographs, or real photos without a free licence and a credit (our images are labelled AI illustrations or credited file photos).
- Personalisation, accounts, comments or a paywall.
- News that isn't about AI.
- Fiction presented as news. Our one kind of fiction is a Special (Halloween, April Fools): a front page shown over the homepage for one day, the page labelled "Fiction" at its foot, written by AI desks. Specials never name real people, never link anywhere, always turn on something impossible, and are not part of the record: no story pages, feed items or archive entries.
How it’s made
Each night a pipeline of AI models reads the sources listed on Sources, drafts stories, and has other AI models check every claim against the source text. A story that fails the checks is rewritten down or dropped. Nothing is read by a human before publication. Run outcomes are public on Status, and fixes are logged on Corrections.
Models in use
What the Sunday, November 22, 2026 edition actually used, by role. This changes from day to day.
| Role | Model | Provider |
|---|---|---|
| fact-checker | Qwen/Qwen3-235B-A22B-Instruct-2507@deepinfra | hf |
| fact-checker | google/gemma-4-31B-it@deepinfra | hf |
| fact-checker | gpt-6-luna | openai |
| fact-checker | jev-1.13.0 | typesafe |
| writer | claude-opus-5-5 | anthropic |
| writer | Qwen/Qwen3-235B-A22B-Instruct-2507@deepinfra | hf |