Method
How ScrollSci works
Every ScrollSci card is built the same way: one item is pulled from a public feed, its own text is used as the only grounding, and three model calls turn it into a 30-second summary, a “why it matters” line, and a brief for an illustration. The card is stored once and shown to every reader who picked that field. Nothing is generated per visitor.
What happens between a source feed and a card in your feed?
Five steps, in order. The important one is the second: the model never starts from a blank page, only from text that was fetched first.
-
1
Fetch
A recent item is taken from the chosen field's public feed — an arXiv preprint abstract, a NASA newsroom release, a Hacker News front-page post, a Wikipedia “on this day” entry, or a Project Gutenberg release. All seven feeds are free and open; none of them require a key.
-
2
Ground
The item's own text becomes the only material the summary may draw on. Where the feed gives no text of its own — most Hacker News front-page posts are bare links — the linked page's own description is fetched and used instead. If there is still nothing substantive, that is recorded explicitly rather than papered over.
-
3
Write
Three separate calls go to Google's Gemini models: one for the 30-second summary, one for the “why it matters” line, and one for a short brief describing an illustration. Each is instructed to stay within the fetched source and to hedge where the source is thin rather than fill the gap.
-
4
Illustrate
The image brief goes to an image generator, which returns one editorial-style illustration. It is decoration and mood, not evidence — see the limits below.
-
5
Publish
The finished card is written to the shared pool, tagged with its field, and carries the source name and link. From then on it is served to every reader who selected that field.
Which feeds does each field draw from?
Seven fields, seven feeds, no others. Every one is publicly readable, so any card can be traced back to something you can open yourself.
| Field | Feed | Material | Caveat worth knowing |
|---|---|---|---|
| Biotech | arXiv q-bio | Preprint abstracts | Not peer reviewed |
| Aerospace | NASA newsroom RSS | Official press releases | Written to promote the programme |
| Computer Science | Hacker News API | Front-page posts | Community-voted, not editorially selected |
| Mathematics | arXiv math | Preprint abstracts | Not peer reviewed |
| Physics | arXiv physics | Preprint abstracts | Not peer reviewed |
| History | Wikipedia “On this day” | Dated events | Community-edited encyclopedia |
| Literature | Project Gutenberg | Newly released public-domain books | Release date is not publication date |
What happens when a source has almost nothing to summarise?
This is the case where summarisers usually start inventing, so it gets handled explicitly.
Most Hacker News front-page items are link posts: the API returns a title and a URL and no body text at all. A summariser handed only a headline will happily produce five confident sentences about a paper it has never seen.
ScrollSci fetches the linked page's own meta description first and uses that as real grounding. When even that comes back empty, the pipeline records an explicit “no information available” marker instead of a plausible-looking blank, and the writing prompts are built to hedge honestly on thin material rather than invent specifics.
The result is that some cards are noticeably more tentative than others. That is the system working, not failing.
Read this part
What can ScrollSci still get wrong?
Grounding a summary in a real source removes one failure mode and leaves several. These are the ones worth knowing before you repeat something you read here.
A faithful summary of a wrong source is still wrong
Preprints have not been reviewed and some do not survive review. Press releases describe results the way the organisation that funded them would like them described. Accuracy to the source is the guarantee; accuracy to reality is not.
Compression removes caveats first
Thirty seconds of reading cannot hold a sample size, a confidence interval and a limitations section. What survives compression is the claim, which is the part that most needs the caveats.
“Why it matters” is an interpretation
That line is written by a model reasoning about significance. It is the least source-bound thing on the card and the most likely to overstate. Read it as framing, not as a finding.
The illustrations are not evidence
Each image is generated from a short text brief. It does not depict real apparatus, real people, or real data, and should never be read as a photograph or a figure.
A feed is not a field
arXiv's physics listing, NASA's press office and the Hacker News front page each have their own slant on what is worth posting. Reading them is not the same as surveying the discipline, and no card should be mistaken for a consensus view.
Review is reactive
Flagged cards are checked by a person and corrected or removed, but that happens after a reader reports it. There is no editor reading every card before it publishes.
The honest summary. ScrollSci is a good way to find out what is happening and a bad way to settle an argument. If a card matters to you, open the source it names and read that instead.
How do I report a card that is wrong?
Use the flag button on the card itself. It opens a comment box — say what is wrong, and the card goes into a review queue to be corrected or removed. That is the fastest route, because the report arrives attached to the exact card.
For anything that isn't about one specific card, email [email protected].
See it working
Pick your fields and scroll. It takes about a minute to set up, and it's free.