Powered by Jev from TypeSafe
How ParhLo checks every story.
Jev is a decision model. It doesn't write anything. It answers precise questions about each article, with a probability for every answer. The language model writes your paper; Jev decides what goes in it and checks what it says.
One article, four decisions.
An illustrative example of what Jev answers for every article.
Incoming article
State Bank keeps interest rate unchanged at 11pc
The central bank held its policy rate for a third meeting, citing easing inflation and a stable rupee.
Business Recorder
After the summary is written
Yes / no“The State Bank kept its policy rate at 11% for a third meeting, pointing to easing inflation and a stable rupee.”
Does this summary say anything the article doesn't?
✓ Checked against source · 92%
Five places Jev works in your paper.
Which topic is this?
Jev picks one of 17 topics for every article, with a probability for each. That is how a story lands in your Cricket or Economy section.
17 topics · 1 call per article
Is this real news?
Quizzes, podcast listings, tag pages and promotions are flagged and kept out of your paper.
Hidden when 78% sure it isn't news
How important is it?
Every story is rated on five levels, from filler to top national news. The two biggest stories of the day go to every reader.
5 levels · 2 Top story slots
Is the summary faithful?
After a summary is written, Jev reads it next to the original article and looks for any claim, number, name or date the article doesn't support.
✓ only when under 30% risk
Which story is your question about?
When you ask your paper something, Jev decides which of today's stories it's about, so the answer comes from the right articles.
Routes every question
Measured, not promised.
0%
of summaries judged correctly in our test of real and deliberately falsified summaries. The same accuracy as the language model it replaced.
0 ms
median time to check a summary, against 941 ms for the language model. About two to three times faster.
0
articles sorted, filtered and scored in 30 seconds in one nightly run. 139 of them were flagged as not real news.
Measured on ParhLo's production system. The summary test used 16 real and deliberately falsified summaries.
Jev decides
- Typed answers: a choice, a yes or no, or a score
- A probability with every answer
- Fast enough to run on every article, every night
The language model writes
- Summaries, written only from the original article
- Headlines, the “why this story” line and the layout
- Answers when you ask your paper a question
See it on your own paper.
Hover any “Checked against source” badge to see Jev's confidence for that story.