The machinery

How Astro Scan Works

This site publishes curated safety intelligence. The gathering and sorting is done by a small set of automated agents. The decision to publish is not.

Nothing is published without a human reading it first. Every item is reviewed and approved by a person before it appears. There is no automatic publishing path, and there is no plan to add one.

Where the material comes from

The agents monitor 14 sources across the four subjects: official safety bodies, national regulators, defence science organisations, and academic literature. The list starts deliberately small so that the scoring can be judged before it is scaled up.

The four agents

gathers

Collector

Plain, deterministic software, not an AI model. Several times it checks a fixed list of official and academic sources, reads what is new, and sets it aside. It respects each site’s rules, identifies itself honestly, and never overloads a source. A source that goes quiet is noted, not hidden.

sorts and scores

Classifier

For each new item, a fast AI model decides which of the four subjects it belongs to and scores how relevant it is from 0 to 10. It must write one plain sentence explaining the score. That sentence appears under every published item, so you can judge the reasoning, not just trust it. Anything scoring below 5 is set aside and never queued.

summarises

Summariser

A stronger AI model writes a short, factual, two to three sentence summary and links out to the original. It never republishes the source text. The point of the site is to help you decide what is worth your time, then send you to the primary source.

assembles the weekly digest

Digest editor

Once a week, an AI model groups the week’s published items by theme and drafts the skeleton of the digest. The editorial framing, the part that gives the digest a point of view, is written by a person, not the model.

The path from source to page

Each day the collector fetches new material and the classifier scores it. Items that clear the threshold are summarised and placed in a review queue. A person reads that queue and marks the ones worth publishing. Only then does the site rebuild and update. The whole loop runs on scheduled automation with a single human checkpoint in the middle, and no server of its own.

→ see the current status and the last run