How the research radar works
The full method, including what it refuses to publish.
16 sources are watched continuously. Most of what they publish never becomes an article here — the interesting output of this system is as much what it declines to write as what it writes.
- 1
Detection
YouTube channels push to us over WebSub: the hub sends a signed notification within seconds of upload. Newsletters and blogs are polled with conditional requests, so a check that finds nothing new costs the publisher a few bytes. Pages with no feed are snapshotted and diffed.
Polling stays on even for push sources. Hubs drop deliveries, and a missed notification with nothing behind it means a story is never covered at all.
- 2
Clustering
Items covering the same development are grouped into one story. The strongest signal is a shared primary source — two reporters linking the same paper are certainly covering the same thing. Failing that, shared model and organisation names plus text similarity, weighted so the vocabulary every AI newsletter shares does not make everything look related.
Two channels belonging to the same publisher count once, not twice. They are one voice, not independent confirmation.
- 3
Reading the primary material
For a video, the pipeline fetches the transcript and chapters, then pulls the links out of the description — where these creators put the papers, repositories and model cards they are covering — while discarding sponsor and social links.
Each remaining link is resolved against the source of record: the arXiv API, the GitHub API, the Hugging Face API. A link that will not resolve is not a citation, and the writer is not permitted to use it.
- 4
The publish gate
A story is only written up once two independent publishers have covered it, or one trusted publisher has covered it with a primary source that resolved. A single listing or a lone low-trust mention is not an event, and waiting costs nothing.
- 5
Writing
A language model drafts the article under a fixed editorial policy. The rules that matter: never state a fact that is not traceable to the supplied evidence, never invent a citation, attribute anything only one reporter said to that reporter by name, and separate what the primary source claims from what the coverage extrapolated.
Charts are re-plotted from reported numbers rather than screenshotted, and diagrams are drawn from the mechanism the source describes. Original figures are embedded only where the licence permits it.
- 6
Validation and review
Every draft is checked mechanically before it can become a file: citations must exist in the evidence bundle and must have resolved, figures must be licensed, claims marked as documented must point at a real source. A draft that fails is rejected rather than patched.
What survives is opened as a pull request and reviewed by Alfred Kolakkal. Nothing publishes without that.
- 7
Revision
When a later source adds something to a story that is already published, the article is revised in place — integrated into the section it bears on, not appended as an update note. The URL and headline stay put so inbound links keep working.
If new evidence contradicts something already published, the article says so explicitly and in the past tense. Silently correcting a published error is the one thing that would make this untrustworthy.
What it will not publish
- Stories resting on a single low-trust item with no primary source.
- Anything where every claim would be secondary reporting.
- Product launches with no technical substance, where the honest article would be a press release with extra steps.
- Revisions that add nothing — the source is recorded as corroboration and the article is left alone.
Questions
Are these articles written by AI?
The drafts are. Sources are detected automatically, the primary papers and repositories are fetched and read, and a language model writes the draft under a fixed editorial policy. Every article is then reviewed by a human before it publishes, and each one says so on the page.
How is this different from summarising the videos?
The reporters are how a story is found; the article is written from what they were reporting on. When a video links a paper, the pipeline fetches the paper from arXiv, reads the abstract and metadata, and the article cites the paper. A claim that exists only because a reporter said it is attributed to them by name.
What happens when several sources cover the same thing?
They are grouped into one story and produce one article. Later coverage revises that article in place when it adds something — a new primary source, a contradicting number, a detail the first report missed. Coverage that merely restates what is already there is recorded as corroboration and changes nothing.
What stops it from publishing something that is wrong?
Several things. A story needs two independent publishers, or one trusted publisher with a resolvable primary source, before anything is written. The model may only cite sources the pipeline actually fetched — an invented citation fails validation and the draft is rejected. Every claim is tagged with the evidence behind it. And a human reviews the result.
Are figures from papers reproduced here?
Only when the licence permits it, which is rarely — arXiv’s default terms grant arXiv a distribution licence, not us a reuse licence. The default is to re-plot the reported numbers as an original chart, which is lawful under any licence because facts are not copyrightable, and to draw original diagrams of mechanisms described in the text.
The editorial policy is version-controlled at pipeline/radar/generate/constitution.md, and each report ships the exact evidence bundle it was written from. See what it watches.