Methodology

What Mantle does and does not claim

Mantle shows who is covering a story, how each part of the political spectrum frames it, and what is missing from a given news diet.

Mantle does not decide which account is true. It does not rate individual articles as biased or unbiased, and it does not endorse any outlet. The labels exist so you can see the shape of the coverage and judge for yourself.

There are two exceptions, each held to a written rule and described below. Mantle marks a claim when the coverage reports evidence against it, or a fact established for years contradicts it. And it marks wording that attacks people for who they are. Both marks are about the claim or the words and whoever is behind them, never about an outlet that reports them. Everything else is described, not judged.

The five-point spectrum

Every outlet on Mantle has a place on a five-point spectrum.

The same test is applied on both sides. Left and Right are for outlets whose coverage leans one way while mostly reporting in the conventional manner. Far Left and Far Right are for outlets whose coverage consistently and strongly favours one side, in the stories they choose, how they frame them and the words they use. "Far" describes how strong and how consistent the lean is. It is not a claim that the outlet is extreme or unreliable.

Mantle's five points line up with the five points other published ratings use. An outlet that independent raters place at the outer point of their scale is Far Left or Far Right here.

Outlet ratings

A rating is Mantle's own editorial judgement about an outlet's news coverage, not its opinion section. It is formed from four inputs:

1. Story selection: which events the outlet chooses to cover and how prominently.
2. Framing: the terms, emphasis and omissions that recur across its coverage.
3. Sourcing: who is quoted and who is not.
4. Public consensus: independent published ratings are consulted as one input. Mantle does not copy them, license them, or present them as its own.

Each rating carries a confidence level (high, medium or low) and a written note. The full rating history of every outlet is public on its page.

Outlets outside the United States are rated on the same spectrum relative to American politics, because that is the frame most Mantle readers compare against. This is stated on the page of every such outlet.

A rating describes an outlet. It never describes a single article, and it never says who is right.

The three sides

To compare how a story is told, the five ratings are grouped into three sides: the left (Far Left and Left), the center, and the right (Right and Far Right). Compare Views, framing notes and Hidden Echoes all work with the three sides. Every outlet still shows its own five-point rating.

How stories are grouped

Mantle reads the public feeds of the outlets it follows every few minutes. Articles about the same event are grouped into one story in three steps.

1. Each article's headline and opening are turned into a numerical fingerprint of their meaning.
2. Articles with very similar fingerprints that name the same people, places and organizations are grouped automatically.
3. When the match is uncertain, a language model reads both and decides whether they describe the same event.

Once an hour a second pass looks for stories that were split in two and merges them. Roundups, live blogs and newsletters are left out, because they resemble every story a little and belong to none.

A story appears on Mantle once three outlets have reported it as news. One or two reports are too few to compare how a story is told, and a pair of articles is where a mistaken grouping is most likely. Opinion pieces do not count towards the three.

How stories are chosen

A story is published as soon as it has the coverage described above, and its analysis is written again as the coverage grows: when 4, 8, 16 and 32 outlets have reported it, when a side that was missing starts to cover it, and at most every six hours while new articles keep arriving. Between analyses, the outlet counts and the coverage spectrum keep changing as new articles arrive.

Mantle leaves out sports and entertainment. Before a story is analysed, a language model sorts it by one rule: a story about games, athletes, teams, celebrities, film, music, television or awards is left out, unless it concerns government, law, politics, public money or public safety.

Reports published by more than one outlet

Outlets often publish reports they did not write. Many carry the reports of news agencies such as the Associated Press and Reuters, and some carry the reports of another outlet. One agency report on three websites is one account of the event, not three.

A newsroom counts once. Towards the three outlets a story needs, the reporting of one newsroom counts once, however many outlets publish it. A story covered by one agency report on three websites and by one other outlet comes from two newsrooms, and it does not appear on Mantle until a third reports it.

Mantle treats two articles from different outlets as the work of one newsroom when any of these is true:

1. The credit says so. Both bylines or datelines name the same news agency, or one names the other outlet.
2. The same writer. Both bylines name the same person.
3. The same text. Half or more of their sentences are identical.
4. The same words. Their headlines and openings are nearly identical. This test also works for articles Mantle is not allowed to read in full.

A newsroom can write more than one report on a story, and each is shown. Two articles are the same report when their text or their words are the same, or when they come from one newsroom under one headline. On the story page a report is shown in full once, under the outlet that wrote it or, for an agency, the outlet that published it first. Every other outlet that published it still has its own card, with its own headline and a link to its own page. The card says whose report it is.

The coverage spectrum still counts every outlet that published the story. An outlet that runs an agency report has chosen to cover the story, and its readers have seen it. If Mantle left that outlet out, it could call a side silent on a story its outlets had published. The story page says how many outlets published the story and how many newsrooms reported it.

The analysis reads every report once. A fact found only in the reports of one agency is a fact from one source, whichever outlets published it.

Mantle does not rate news agencies yet, because it cannot follow them directly. An agency report is shown under the rating of the outlet that published it, and that rating describes the outlet, not the agency.

The coverage spectrum

For each story, Mantle counts the distinct outlets at each of the five points, not articles. An outlet that publishes six updates counts once. A point with no outlets is still shown, because the gap is part of the picture.

Opinion pieces are labelled and left out of the count.

Blindspots

A blindspot, which Mantle calls a Hidden Echo, is a story that one side of the spectrum covers far less than expected.

Some parts of the spectrum publish far more than others, so raw counts would mislabel almost everything. Mantle instead compares two numbers for each of the three sides:

• the observed share: the side's share of the outlets covering this story
• the expected share: the side's share of all outlets that published anything in the last 24 hours

A side is flagged when at least five outlets cover the story and the side's observed share is a quarter or less of its expected share. It is also flagged when the side is entirely absent although at least one or two of its outlets would be expected. Every flag shows the arithmetic in plain language.

Neutral headlines and summaries

The headline and summary at the top of each story are written by a language model, Claude, made by Anthropic. It follows editorial rules written by Mantle:

• Describe, do not evaluate.
• Use the most specific neutral noun: a name, a role, "suspect", "protesters", "lawmakers".
• Attribute every contested claim to whoever made it.
• Include only facts that at least two outlets report, or that are clearly attributed.
• Write about the event, not about who has covered it. The coverage spectrum shows that, from counts.
• When coverage from only one side is at hand, do not invent what the other sides would say.
• When only one side reports a story, say which outlets report it and attribute every statement of fact to them. Such a story is never written in Mantle's own voice.

Every headline and summary is then checked against Mantle's own list of loaded words. If one is found, the text is rewritten. If the rewrite still contains one, the story is held back instead of published.

The model, the version of the rules and the time of writing are recorded for every story.

Key points

Under the summary, each story lists the points a reader needs. They are written by the same model under the same rules, and each one shows who reports it: the ratings of the outlets behind it, from Far Left to Far Right, with the outlets' names one tap away.

A short label says how widely a point is reported. Across the spectrum means at least two newsrooms on every side covering the story report it. One side only means every outlet behind it is on one side. One outlet means a single outlet reports it. The label describes the reporting. It does not say the point is true or false: a point that one side alone reports can be accurate and simply not picked up elsewhere.

Framing analysis

For each side that has covered a story, Mantle describes how that side's headlines present the event, what its coverage puts first, and which facts reported elsewhere it does not mention. When the outlets at the far end of a side tell the story differently from the mainstream outlets on the same side, the note says so.

Framing notes are descriptive. They are written by the same model under the same rules, and every side is held to the same standard. An omission is claimed only when it can be checked against the text Mantle has.

Claims that do not hold up

A story page can carry a warning above its summary, under "Read with care". It appears on few stories. There are two kinds, and nothing else is marked.

• Disputed claim. A claim of fact that the coverage gives solid reason to doubt: a record, footage, published data or an official finding reported in the coverage contradicts it, or at least two newsrooms report that it is false or was made without evidence. A boast counts when it asserts something that can be checked. An opponent's denial alone is not enough, and neither is one outlet's opinion.
• Contradicts established facts. A claim of fact that directly contradicts something established and documented for years by the body whose job it is to establish it: certified election results, final court judgments, official statistics, or the stated consensus of the scientific and medical bodies in the field. It is never used for the events of the last two years, where the model may not know what has since been established.

Opinion, values, forecasts, vague praise or abuse that asserts nothing checkable, and contested readings of the evidence are never marked. They appear under "Where accounts differ" when the coverage disagrees.

A warning is about the claim and whoever makes it. It names the speaker, gives their words as reported, and says what disputes them. The outlets that report the claim are listed as reporting it: an outlet that quotes a false claim accurately has done its job. Only when an outlet states the claim itself, with no speaker, does the warning say so. The same rule applies to every speaker, every outlet and every side.

A warning is proposed by the model that writes the analysis, and is published only if a second, separate check upholds it against these definitions. When that check cannot run, no warning is published. A story without a warning has not been fact-checked: Mantle marks only what meets one of the two rules.

Charged language

Mantle highlights words and phrases that carry a judgement beyond the facts. Every highlight shows the category, why the wording is loaded, and a neutral alternative.

Highlights come from two sources.

1. Mantle's word list. A published list of loaded words and phrases. A match is highlighted on sight only when the term is unambiguous and clearly loaded.
2. Reading in context. The language model reads the headline and opening of each article and flags loaded wording where a neutral one was available. Words from the list that depend on context are highlighted only when this reading confirms them.

Accurate strong words are not flagged: an attack that killed people is "deadly". Neither is the word outlets across the spectrum use for the same thing, such as a government "shutdown". Words inside a quotation are marked as quoted, because the outlet is reporting someone else's language.

Every word the language model flags is checked a second time, against the same definitions, before it is shown. The check removes a flag on an accurate word, on the word every side uses, on ordinary writing that judges no one, or on someone else's words taken for the outlet's own. A highlight accuses an outlet of a choice, so Mantle would rather miss a loaded word than flag a fair one.

The word list is built so that terms used mainly on the left and terms used mainly on the right appear in comparable numbers. The current count is shown on the language guide.

Hateful rhetoric

A few of the highlighted words are marked Hateful rhetoric: wording that attacks or dehumanizes a person or a group because of race, ethnicity, national origin, religion, sex, gender identity, sexual orientation or disability. It takes one of these forms:

• likening them to animals, vermin, disease or an infestation;
• a slur;
• saying the group is criminal, dangerous or inferior by nature;
• repeating a conspiracy claim that the group secretly controls others or plots against them;
• attacking one person through their group: mocking their name, faith or origin, or calling them a terrorist, a criminal or a foreigner because of it;
• attaching the group's name to a crime, so that the group is tied to the act;
• approving of violence against them.

Harsh words about politicians, parties, governments, institutions and ideas that do not rest on who the person is are not marked this way. Neither is a term of legal status such as "illegal alien". Those are loaded language and are highlighted in their own categories. So is naming a person's religion, race or origin where it has no bearing on what happened, and tying a person to a feared group through a link the wording does not spell out.

The mark is about the words and whoever said them. When an outlet quotes someone, the page names the speaker and says the outlet is reporting their words; only wording in an outlet's own voice is put to the outlet. The mark passes the same second check as a claim warning, and the same rule applies to every speaker and every outlet. A story with such wording says so above its summary, before the wording itself is read.

How Mantle reads and shows articles

Mantle finds articles through the feeds that outlets publish for that purpose. Before it reads a feed or an article page it checks the site's robots.txt, and it does not read what the site rules out. When an outlet refuses automated reading of its pages, Mantle does not try to get around the refusal. It works from what the outlet publishes in its feed: the headline and the description, and the text of the article where the outlet includes it there.

Mantle shows an article's headline, a short excerpt, the outlet's name and a link to the original. Full articles are never reproduced. Articles belong to their publishers, and the point of Mantle is to send you to them better informed.

How Mantle chooses images

Each story shows one image, chosen from the images the outlets covering it published with their own articles. An outlet's lead image is part of how it frames a story, so Mantle never takes one outlet's image by default. It picks the image that shows the story most accurately.

1. Candidates. Mantle looks at up to ten of the story's images. It leaves out images from opinion pieces, outlet logos and placeholders, images from sites whose robots.txt rules them out, and images too small or too large to use.
2. What each image shows. A language model describes each image against a fixed checklist. It is told not to judge bias, and never to name or identify a person.
3. Published rules decide. The description becomes a score. Images that show the story's subject directly score highest. Photographs, graphics, charts and the logo of an organization in the story are all welcome.
   • Never shown:
     • an image that is mostly text;
     • an image with branding a channel or outlet added on top;
     • graphic injury;
     • a mugshot or someone in custody;
     • a child as the main subject;
     • an image that does not match the story.
   • Points off: anger, shouting or aggression; violence; an unflattering angle; a composite or split screen; poor quality; a generic stand-in. Grief, tears, fear and smiles cost nothing: they are often what the story is about.
   • Points added: when outlets on different sides of the spectrum used the same photograph.
4. Close calls. When two images score about the same, the model compares them side by side and picks the one that shows the story most specifically. A story keeps its image unless a clearly better one arrives.
5. The most-covered figure. On a story about the week's most-covered public figure, Mantle shows no photograph of a person, so the same face does not fill the front page. The rule applies to whoever is named in more than one in ten of the week's stories.

Mantle does not balance images by side. If one side's outlets use more emotive images, theirs are chosen less often. That is the rule working, not a bias. When no image passes, the story shows none.

Every image is credited to the outlet whose article carried it, and links to that article. Mantle stores a resized copy, so pages load quickly and your browser does not contact the outlet. Images belong to their outlets and photographers. Anyone can ask for an image to be removed, and it is taken down the same day.

Your reading history (My Mix)

My Mix shows how the articles you open are spread across the spectrum. The list of what you have read is kept on your own device, and the numbers are worked out there. Mantle does not receive it. You can delete it at any time from the My Mix page.

Only articles you open at an outlet count towards your balance. Reading Mantle's own summary of a story is noted, but it has no rating. Mantle does not describe your balance until you have opened at least five articles in a week, because fewer than that says little.

Known limitations

• The source list is not the whole media landscape. Wire services without public feeds, such as AP and Reuters, are missing until a licensed feed is added.
• Some outlets do not allow automated reading of their article pages. For those, Mantle works from what the outlet publishes in its feed. Where that is a headline and a description only, loaded language deeper in the article is not seen.
• An image's description can be wrong. The model sometimes misreads what an image shows, and the rules then choose with a mistake in hand. Every story image can be reported from its page.
• Some outlets close their feeds to automated reading. Mantle does not read those feeds. The outlet keeps its rating on Mantle, and its coverage is missing from stories until its rules change.
• Ratings are judgements, not measurements. They are published with confidence levels for that reason.
• Grouping can be wrong. It sometimes merges related but distinct events, or splits one event in two.
• A warning can be wrong, and most false claims carry none. A claim is marked only when the coverage Mantle has read reports evidence against it, or an established fact contradicts it, and the model's knowledge of established facts has an end date. Every story has a link to report a mistake.
• Language models can be wrong. Each story says that its headline and summary are written by Mantle's AI, the rules it follows are described on this page, and every story has a link to report a mistake.
• Sorting out sports and entertainment can be wrong in either direction. A news story sorted out by mistake is not analysed.
• News agencies are seen only through the outlets that publish their reports. Mantle cannot follow the Associated Press or Reuters directly, and it does not rate them yet.
• Some copies go unnoticed. An outlet that rewrites an agency report under its own name, or that does not allow its pages to be read and names no source, counts as an outlet of its own. Mantle would sooner count one outlet too many than fold two newsrooms into one.
• Two writers can share a name. Mantle reads a writer's name as it is printed. Two people with the same name, writing about the same story for two outlets, would be taken for one newsroom.