How this deck works, in full

What the deck is

A curated daily brief over other people's published feeds. We read what publishers and creators put out for machines to read, choose what seems to matter, write our own short note about it, and link you to the source. We are not a news outlet, we republish nobody's work, and the writing here is ours alone.

The evidence tiers, which are an argument

Every item carries a coloured tier. The colour encodes what kind of claim the item is, and the tier meaning never rides on colour alone: each chip is also a word. The colours are the same ones the rest of this family of sites uses for evidence, so a reader who knows one knows all of them.

Primary	The organisation said it itself, or this is the paper
Reported	An established outlet covering it
Commentary	Analysis, opinion, or community attention
Unverified	Unattributed, rumoured, or contradicted

A tier says what kind of source something is. It does not say the claim is true. A primary announcement can be wrong and a piece of commentary can be right; the tier tells you where to look to check.

The permission tiers, which are a law

Separately from what kind of source a publisher is, we record what we are permitted to take. This is decided by the publisher's own published signals, chiefly its robots.txt, and it is enforced in one function that every item passes through, with a test that hands that function full text and proves it comes back empty for a restricted source.

A	Open	Explicitly open terms, or no restriction on automated readers at all. We may show the headline, the link, and the feed's own summary in full.
B	Feed open, crawlers welcome	The publisher's robots file admits automated readers. We show the headline, the link, and a short extract the feed itself offers. Never the full text.
C	Headline and link only	The publisher's robots file asks automated tools to stay out. We honour that: the headline and the link, nothing else, whatever their feed happens to contain.
D	Excluded	The publisher blocks even the bots whose only job is to cite and send traffic. A publisher refusing the referral is asking to be left out.

Permission tiers are shown as plain letters on the sources page and never in the evidence colours, because they are not a claim about evidence.

What we never do, and how you can check

• We never fetch an article page. Feeds only. Nothing is crawled.
• We never pass a publisher's text through a model to make a summary. Our notes are editorial framing written from the headline, the source pattern and the tier, and the page says so.
• We never embed a video. Every video is a link, so the view, the session and the subscribe land with the creator. This also means no YouTube API is used at all, so no API agreement governs this site.
• We run one script, and it is ours. _shared/constellation.js draws the moving starfield, and it is the only one: no third-party script, no tag manager, no analytics, no cookie, nothing that reports anywhere. Open your browser's developer tools on any page and you will find exactly that one file, zero cookies, and image requests only to the hosts the privacy page lists, which is itself read off the built site. This site ran no script until 20 September 2026; the change was made deliberately for the background and is recorded rather than glossed.

The reader's door

A story whose link does not open for a reader does not go on the page. Before anything is clustered, every story link from the article layer is checked against the publisher's own machine-readable access declaration: the isAccessibleForFree field and the article:content_tier tag, which publishers emit deliberately so that a machine can tell a free article from a paid one. A paywall, a registration form or a dead link removes the story.

What the door does not cover, stated plainly. Video links go to YouTube and are not probed, and neither are the publisher home pages listed on the sources page, which are not articles. YouTube is free to watch, but that is a reason rather than a measurement, and this deck prefers to say which of the two it has. The reason it stays a reason: that endpoint rate-limits per IP and answers a tripped limit with a 404, so a few hundred probes to confirm something free would risk the block for no information. Members-only videos exist and are not detectable from a link. Of the story links on this site today, the door has ruled on every article and on none of the videos.

Four rules keep that from doing harm, because the failure mode here deletes good journalism rather than adding bad:

• No article text is read or stored. The whole record kept is a status code, a verdict and the sentence of evidence that produced it.
• robots.txt is obeyed, with this deck's own honest name. We do not wear a browser's identity to get past a block. A publisher who says no gets unknown.
• A 403 is not a paywall. A bot-block and a hard wall are the same picture from here, so both are recorded as blocked and the story is kept.
• Ignorance never excludes. An unjudged link stays on the site and is counted in its own column on the ledger, so the number is visible rather than absorbed.

What the deck labels, and what it refuses to label

Every story is sorted on several axes: whether a piece reports or argues, what kind of event it is, what field it touches, whether the thing has actually happened, where, and who did it. Those labels are how the deck knows what the AI conversation is actually about rather than guessing.

An axis only appears on the site once it has cleared a published accuracy floor of 70%, measured on a sample of stories it was never tuned against. Today one axis of the four has. The others are computed, used internally to choose what to show you, and print nothing at all, because choosing a story with a weak instrument costs a reader nothing while putting a label on it would be a claim.

And the figure is called agreement, not accuracy, on purpose. The reference labels were made by an AI model reading headlines, and that same model wrote the classifier being scored, so the number is agreement with one judge marking its own work. The ledger carries the full table, the failing grades, and what it would take to turn it into a real accuracy.

Keeping this current

Where a fact about the outside world is checked, it is checked against the newest source available, not against what was true when the page was written. Anything dated on this site carries the date it was checked, and a claim we have not rechecked is attributed to whoever reported it rather than stated as ours.

The two refusals

Two rules are enforced by the build failing rather than by anyone remembering:

• A story with no note of ours cannot lead. It goes to the plain index with its headline and link, which is all we are entitled to show without adding something.
• A creator with no write-up of ours gets no page. The difference between our page about a creator and a copy of their channel is the writing, so the writing is mandatory.

Ordering

Every list on the site that claims to be chronological is checked to be chronological on every build, by reading the ages the page itself prints. The leads and the index run newest first; the creators run by their newest upload. This is stated because it was wrong three times before it was checked.

Times

Every time is shown twice: how long ago, to the minute, and the absolute time in 12-hour US Central. The store keeps UTC; the conversion happens once at display and is tested at the daylight-saving transition instants and against the operating system's own clock.

The schedule

The deck refreshes twice a day at a random time inside two five-hour windows, 05:00 to 10:00 and 17:00 to 22:00 Central, never less than seven hours apart. The windows are five hours wide because that is the widest the seven-hour rule permits; the arithmetic and a 20,000-day simulation are in the build notes.

Images

Where a publisher's feed offers a picture and the permission tier allows it, we show that picture. Where none is offered, the story gets a generated diagram in its tier colour, labelled as such, and never an invented photograph: a photorealistic image beside a real headline reads as a depiction of the event, and no caption reliably undoes that.

Corrections and removal

Errors are recorded on the corrections page the day they are fixed, with what was wrong, what changed and how it happened. The original text is described, not erased. Any publisher or creator who would rather not appear here is removed within 24 hours of asking, no reason required. Details for publishers are on their own page.

Disclosure

The editorial notes and creator write-ups are drafted by an AI collaborator and published under human editorial responsibility. Headlines, links, dates, thumbnails and source names come untouched from each publisher's own feed.

The archive

Each day's leads are kept, with the note as it read that day, on the archive. Only stories we wrote about are kept: the headlines we merely linked belong to their publishers, and the record here is of our own work.