How this works

Each channel gathers a week of stories and publishes them here. What differs between channels is how much a language model is trusted to do — and for most of them, the answer is very little.

Where the facts come from

Release Radar uses no model at all. Versions and end-of-life dates come from the GitHub Releases API and endoflife.date, so every number is checkable at the source. Security takes its exploited-vulnerability half straight from CISA's catalog, with severity from NVD.

Engineering Reading and arXiv Picks let a model choose and summarise, but never write a link or a title — those come from the feed or the paper record, so an article that does not exist cannot be invented.

Only the three news channels let a model find stories. Even there, every link is one it actually retrieved rather than composed, filtered against a list of publications chosen in advance; where the stated source disagrees with the link, the link wins.

What you will not see

Padding. If a story has nothing non-obvious to say about why it matters, it says nothing instead of filling the space.

The figures behind all of this — what supplies each channel's facts, which publications appear most, how many stories are withheld and why, and how many published links still resolve — are on Where this comes from, computed from the archive each time the site is built.

Channels