How the niche pages are counted and checked.

The sample

Hazel runs a crawl every day across 211 niches on Reddit, YouTube and Facebook, using a fixed set of search queries per niche that does not change between runs. Each conversation the crawl turns up is one row in the sample. A page only publishes once its niche and platform hold at least 15 rows, and it never renders on less.

The score

Each conversation is read by the same classifier that scores the daily board in the game, a Claude model, and given a signal type and a score from 0 to 100. The types are pain, buying intent, discovery, rival, and visibility. Everything else is marked as not a signal and stays in the count, because the ratio of real asks to everything else is the most useful number on the page.

What is never stored

The sample keeps the community, the signal type, the score, and the date. It does not keep the text of the post or a link to it. We classify, then discard, and we do not quote people who did not ask to be quoted. The live posts, with a reply drafted for each, are only in the game.

The window

A page’s honeycomb and every number in its prose come from the most recent sample. The Monday email a watcher receives counts this week against last week. The quiz counts the last 30 days. Each surface says which window it is using.

Meadowrank’s own model writes each page for its niche and platform, from the rows behind that page. It is the same family of model that scores the board, working from the sample and a house brief, not from open-ended invention. No number in the prose is typed by hand. Every count, ratio and number word is computed from the same sample the honeycomb renders, so a re-count can never leave a sentence contradicting the grid. Named communities must appear in that page’s own rows. Facebook groups are described and never named, because pointing strangers at a small local group is how its members get banned.

The date on each page is the day its sample was last counted, which is the day its numbers last moved. It is not touched otherwise.

Every page is signed off before it publishes, and signed off again when its numbers move enough to change a sentence. Right now the person doing that is the founder. Sign-off means reading the page against its own rows and spot-checking the claims, not reading every page word by word. A page that gets something wrong comes down, and the correction goes live with the next count. Other reviewers will be listed here as they join.