About & methodology
Evals Review helps AI builders find evals that actually catch failures, and hear from people who have used them. It is an independent project run by Siddharth Mishra in India.
How listings are chosen
- The eval must have a public original source: a paper, code repository, dataset or the creator's official page.
- We open every source link before listing it and remove listings whose source disappears.
- It must test something useful to people building with AI: coding, agents, support, writing, safety, healthcare, multilingual use, retrieval and so on.
- Summaries are written in our own words. Facts such as dates, licenses and dataset sizes are taken only from the source. If we haven't checked something, the page says “Not checked yet” instead of guessing.
- Creators are always credited, and every listing links back to the original.
Editorial independence
We are not affiliated with the labs, companies or researchers whose evals we list. Nobody pays to be listed, and nobody can pay to change an editor note, a rating or a review.
Sponsored placements
- Sponsors can pay for a clearly labelled placement, for example a featured spot in a category or the newsletter.
- Sponsored items are always marked “Sponsored”.
- Sponsorship never affects ratings, reviews, editor notes or which evals we list.
No fake reviews
We never invent ratings or seed reviews. A page with no real reviews says “No reviews yet”. We don't buy reviews or reward people for positive ones.
Review rules and moderation
- Reviewers sign in with GitHub. One review per person per eval.
- Reviews should describe real experience of using the eval: what it caught, how easy it was to run, and whether it's still current.
- Not allowed: reviews of your own eval or a competitor's without saying so, personal attacks, spam, promotion, private or confidential data, and anything illegal.
- We may remove reviews that break these rules. We don't remove reviews just because they are negative.
- Creators can reply or report errors through the Contact page.