Explained

How The Morning Science works

Every two hours 35 news feeds, 0 YouTube channels and 0 podcasts go in. What comes out, once a day, is only what a science reader should have heard about. This is every step in between – and what each one did today.

Today’s edition, in numbers

Stories collected
35 feeds · 25 outlets
137
Pieces of news
after grouping the same news
135
In today's paper
passed the news check (60+)
76
Briefed by Claude
facts, timeline, community
60

The pipeline, step by step

1
Collect
No AI – feeds
Today: 137 stories from 25 outlets
Every feed is read at least once a day, and again whenever the paper is older than two hours. Sources are journals and their news desks (Nature, Science …), science magazines (Quanta, Science News, Physics World …), agencies and labs (NASA, ESA, MIT …), Hacker News, across 8 sections. All sources.
Impact · Nothing is picked by hand: the same open feeds every day, so the paper can't be steered by who pays.
2
Group the same news
Claude Haiku 5.5 · rules
Today: 135 pieces of news · 2 covered by several outlets
- Same article URL → same story, always.
- Claude reads new headlines against the threads of the last three days and groups stories about the same specific event – “React 20 released” from three outlets, not “two React articles”.
- If Claude is unavailable, near-identical headlines are matched by their words.
- Security incidents are merged once more when they share a CVE or Claude sees the same incident.
Impact · One card per piece of news instead of five near-identical headlines – and “3 sources” becomes a signal.

The news check

Claude Haiku 5.5 · rules as fallback

Today: 137 stories rated by Claude · 76 of 135 made the paper

Every story gets a score from 0 to 100 for “should someone who follows science closely have heard about this today?”. Claude doesn’t pick that number. It answers the same yes/no questions for every story, judging the event, not the writing, and the score is the sum: start at 20, add the weight of every “yes”.

Question	If yes
Something was found. Does it report a specific new result, publication, mission event, approval or policy decision?	+35
A notable step. Would researchers in the field call it a notable step rather than incremental?	+15
Matters widely. Does it matter beyond one subfield – to many people, many fields or public decisions?	+15
Primary source. Is the paper, dataset or official statement linked or clearly identified?	+10
Independent voice. Does it quote a researcher who was not involved in the work?	+10
Ongoing story. Is it part of an ongoing story people are following?	+5
Explainer or listicle. Is it an explainer, a listicle or a retelling of known science?	−30
Press release only. Is it only an institution's press release, without reporting or a primary source?	−15
Pseudoscience. Does it present fringe claims as findings?	−50
Advice. Is it lifestyle or health advice rather than reporting?	−30
Off topic. Is it outside science and research?	−40

0–29 · Not news

tutorials, “I built X”, listicles, marketing

30–59 · Interesting, not news

essays, niche releases, trending repos

60–79 · News

notable releases, launches, ecosystem events

80–100 · Must know

major releases, serious vulnerabilities, outages, licence changes

▲ The paper starts at 60

Real-world signals push it up: +15 when two independent outlets cover it, +10 more at three, +10/+20 for 50/200+ comments on Hacker News. The best story of a thread decides for all of it. Only 60+ makes the front page, the sections, the feed and the sitemap. Without Claude, rules stand in: the outlet’s track record, “announces something concrete” (+20) and “reads like a tutorial” (−40).

Impact · 59 pieces didn't make it today – tutorials, listicles and marketing stay out.

4
Read the room
Hacker News · Claude summarises
Today: 0 stories with a science discussion
For active stories the comment threads on Hacker News are fetched. Claude sums up the mood, the recurring themes and the arguments for and against – quotes are copied word for word and attributed.
Impact · You see what science readers actually think – not only what the press release says.
5
The briefing
Claude Haiku 5.5
Today: 60 stories briefed
Claude reads all sources of a story and extracts structured facts: the TL;DR, why it matters, concrete action items, a timeline of what happened when, versions, CVEs and companies involved. It doesn’t write articles and never adds facts the sources don’t contain. The original reporting is always linked.
Impact · A two-minute read gives you the facts, why they matter and what to do – with every source one click away.

Watch & listen

Claude Haiku 5.5 · feeds

Today: 0 of 0 videos · 0 of 0 podcast episodes made the paper

The uploads of 0 science YouTubers and the episodes of 0 podcasts are followed. Here the question isn’t “is it news?” but “is it worth a science reader’s time?” – a deep interview with the people who built something belongs in the paper without a news hook; a tutorial, a sponsored pitch or a celebrity chat doesn’t. Claude answers the same yes/no questions for every video and episode from its title and description, and the score is the sum: start at 20, add the weight of every “yes”. Only 55+ is shown.

Question	If yes
Current topic. Is it about something found, launched or decided recently?	+30
Substance. Does it go into how the result was obtained, what the evidence is and where its limits are?	+25
Widely relevant. Does it concern questions many people or many fields care about?	+15
Researchers speak. Do the people talking know the subject first-hand?	+15
Promotion. Is it mainly promotion?	−30
Pseudoscience. Does it present fringe claims as science?	−50
Off topic. Is it outside science and research?	−35
Filler. Is it filler rather than a real episode or video?	−25

Claude also writes a short brief for each video (0 so far). Episodes play right on the page.

Impact · The newest worthwhile video is on the front page, and each section shows the day's video and episode.

7
The paper
Ranking rules
Today: 76 stories in today's edition
The biggest stories go to the top: news score, number of outlets and discussion size decide. Each section shows its top three on the front page. The same edition feeds the RSS feed, the sitemap and the daily archive.
Impact · The front page is an editor's pick, not an index – everything else is one click away.

Who decides what

Decision	Made by	Without AI
Which sources	People (the source list)	—
Same news?	Claude Haiku 5.5	Matching URLs and headline words
Is it news?	Claude Haiku 5.5 answers the checklist, fixed weights decide	Outlet track record + headline rules
The briefing	Claude Haiku 5.5	No briefing, just the sources
Security tier	Claude Haiku 5.5 answers the checklist, fixed rules decide	Estimated from older data
Video or episode: worth it?	Claude Haiku 5.5 answers the checklist, fixed weights decide	Videos: title keywords · episodes: all shown
Order on the page	Ranking rules	—

A small, fast model is used on purpose: a few hundred short extraction calls a day, each one checkable against its sources.

What the AI has used so far

Every call to Claude is logged with the tokens it read and wrote. These are the real totals, counted call by call; euros are the Claude API list price of each call, converted from US dollars.

Tokens so far: 677.4K; since November 7, 2026
Cost so far: €0.107; 132 calls · API list prices
Today: 677.4K; €0.107 · 132 calls
Read / written: 535.7K / 141.7K; prompts & sources / answers

Briefings & discussions
120 calls
514.3K · €0.069
The news check
8 calls
102.3K · €0.026
Grouping the same news
4 calls
60.8K · €0.012

The ground rules for the AI

✓ Only from the sources: Briefings, checklists and video briefs use only what the linked sources say. Unknown stays empty.
✓ Evidence for every “yes”: Security facts name the outlet that reported them.
✓ Quotes are verbatim: What science readers say is copied word for word and attributed.
✓ No articles: The AI extracts facts – it doesn’t write stories. The reporting stays with the outlets, always linked.
✓ Same rules for everyone: No paid placement. Sources and rules are public on this page and the sources page.