How Rally chooses stories

1. Rally only reads established newsrooms

Every story starts as an RSS feed from a publisher on a fixed list — wire services, public broadcasters, national papers and specialist outlets. Rally never pulls from social media, aggregator farms or sites without a masthead. The full list is on the publishers page.

A run picks a sample of those feeds rather than all of them, so the same few outlets do not dominate every edition.

2. A first pass on language

Headlines carrying the vocabulary of violence, disaster, crime and collapse are rejected outright before anything else looks at them. This is a blunt instrument and we know it: it is fast, it is cheap, and it removes the bulk of the day's hardest news without a model having to read it.

It also has an obvious failure mode. A story about a war ending contains the word "war". Stories like that are caught by this pass and lost. We would rather lose some good news than let through a headline that lands like a punch.

3. A language model scores what is left

What survives is read by a language model and scored against written criteria: whether anything actually improved, who benefits, whether the good is real or merely announced, and whether a reasonable person would finish the story feeling better informed rather than merely soothed.

The model is calibrated with worked examples — real headlines with the scores we think they deserve — so the scale means the same thing from week to week. The result is an average across several scoring questions rather than one number plucked out, and a story has to clear a threshold to be published at all.

That score is internal. We do not show it next to a story, because a number beside a headline invites you to read the number instead of the story.

4. Duplicates are removed

Each candidate is compared against what Rally has recently published. If it is the same event in different words, it is dropped. One good thing should appear once, not six times because six newsrooms covered it.

5. Where AI is used, and where it is not

A language model does the filtering, the categorising, the summary you read on a Rally article page, and the extraction of topics, places and people. It is used to decide what to show you and to describe it.

It does not write the journalism. The reporting is the publisher's, always, and it is linked and credited. Where Rally publishes something of its own and a model had a hand in it, that piece is flagged as containing AI and carries no human byline — we would rather admit it than borrow a person's name for it.

6. People are still in the loop

Every run is reviewed. Anything that should not have made it through can be pulled from the site, and pulled stories stay recoverable rather than being silently destroyed. Readers can flag a story too, and flagged stories are read by a person.

What this cannot do

A positive-news filter is not a view of the world. It is a deliberately partial one, and a diet of only Rally would leave you badly informed about the day. Rally is meant to sit alongside a full news source, not to replace one.

Corrections

Rally does not alter a publisher's words, so a factual error in a story is theirs to correct and the link will take you to their current version. What is ours to fix is a bad summary, a miscategorised story, or something that should never have passed the filter. Write to support@rally.news and we will correct or remove it.