The short version
Find the links
We read public Hacker News stories and comments and keep links to individual public posts. A post needs at least three distinct HN authors citing it before it enters the pool.
Read the post itself
The post text, author, date and media are archived from public embed data. Deleted posts and posts we cannot fetch are dropped, with the reason recorded.
Ask what will last
Three evergreen questions score each post from 0 to 1. Posts scoring below 60% average, or that are news, drama, promotion, launches or jokes, are left out.
Count independent recommendations
Every substantive HN comment on a surviving post is read for stance. Only posts recommended by at least two distinct commenters are published, and each comment is shown with its author.
Rank, then get out of the way
Distinct recommenders come first. Ties break on the configured formula: other citing authors, discussion size, recency, and how many different years the idea kept coming back.
Thresholds
- Evergreen score
- 60%
- Distinct recommenders
- 2+
- Citing authors
- 3+
Average of three model questions about lasting value.
Different HN accounts whose comment recommends the post.
Different HN accounts linking to the post, before scoring.
About the AI step
Evergreen and stance labels come from Amos, a decision model reached through a public API. It never decides what is true or important on its own: it answers fixed, narrow questions, and a post still needs independent human recommendations to be published. Public post text and HN comment text are sent to that API, which states that submitted requests may be retained to improve its models.
What this is not
- No submissions, posting, replies or votes. The archive is read-only.
- No engagement ranking. Engagement counts are shown, never used to qualify a post.
- No outrage feed. Negative and neutral citations are visible on detail pages but do not add recommendations.
- No accounts and no tracking scripts. Saved ideas live in your browser.