How we work
Speedrun is made by machine. That means the questions you would normally put to a newsroom - who decided this mattered, where did the facts come from, who checked it, what happens when it's wrong - still need answering. Here are the answers.
What we cover
Tech and AI. One test decides everything: has something actually changed, and does it change something else? A funding round on its own is not a story. A funding round that resets what a category costs to build is.
We are not trying to cover everything that happened. We are trying to cover the things worth having explained.
We also don't chase breaking news. Apple News and X will beat us to the alert every time and we are not competing for it. We publish once there is enough coverage to compose from - later than the first report, and considerably better informed than it.
Where the facts come from
Every speedrun is built from many sources, never one. The number is on screen the whole time you are watching. Open it and they are sorted three ways.
Primary is the thing itself. The filing, the paper, the actual post, the transcript, the benchmark.
Analysis is people arguing about what it means. Named journalists, researchers, practitioners.
Aggregated is wires and roundups, which establish what is broadly agreed.
Where a primary source exists, we read it. Where a story is all analysis and no primary, that is visible in the split.
Who checks it
A person reads every speedrun before it publishes. Every one, without exception.
The job is narrow on purpose. They are checking that each claim is carried by a source, that nothing has been invented, and that the story hasn't quietly acquired an opinion on the way through. They are not there to add a take. If they think the piece is wrong, it doesn't go out.
What we are bad at
Three things, stated before you find them.
We inherit the shape of the coverage. If the entire press writes up a launch from the same briefing, we compose a picture out of that briefing. Many sources is not the same as independent sources, and we cannot manufacture the second.
The models we use were trained on the internet, which is not a neutral place. We can give them rules. We cannot give them a different history.
And we pick what to cover. That is a human judgement, made by three people, and it is the largest editorial decision in the product. Nothing downstream corrects for it.
When we get it wrong
We will. When we do, we fix the speedrun, mark it corrected on its page, and list it below with the date and what changed. We don't quietly re-render.
Think we've got something wrong? support@kerva.com
Corrections
None yet.