How BlockBoard works

Every number on BlockBoard comes from a rule. This page states the rules, the thresholds and the limits, so anyone can check what we say against how we got there.

Last updated October 17, 2026

1. What we read

BlockBoard watches 160 public sources:

• Specialist crypto newsrooms.
• Central banks, regulators and official statistics.
• Exchange announcements, project blogs and governance forums.
• Security alerts and geopolitical news that moves markets.

Each source has its own rhythm. The fastest are checked every one to five minutes, and the slowest once a day. When we download an article page, we first read the site's robots.txt and respect it.

We keep the text of articles on our own servers so we can group and compare them. We never publish it. What appears on BlockBoard is a headline, a link to the original and our own words.

2. From articles to stories

Every few minutes, each new piece is compared with the stories already open. Three things are measured:

• Meaning. A language model that runs on our own server turns each piece into a vector, so two articles that say the same thing in different words end up close together.
• Names. Which companies, coins, people and institutions both pieces mention.
• Wording. Fingerprints that catch copies and near-copies of the same text.

Close enough, the piece joins the story. Far enough, it starts a new one. In between, a second language model is asked one question: is this the same event or not?

Some rules can't be overridden. Two pieces that share no name at all are never merged, and neither are two different kinds of event (a hack and a listing, say).

A story appears on the board only when at least two sources cover it. After six hours without new pieces it is marked quiet, and later it closes. How much later depends on the kind of event: a hack stays open longer than a price report.

3. Who was first

For every story covered by two or more newsrooms, we compare the publication time each newsroom declares. It's a race, and we keep it fair:

• Only newsrooms take part. Official statements are the source, not competitors.
• Syndicated copies don't count.
• A whole race is thrown out if any time is implausible: more than an hour after we first saw the piece, or more than 30 days before.
• A lead of under two minutes is a tie. A lead of more than six hours isn't a race.
• A newsroom needs at least 10 races in the last 30 days to appear in the ranking.

The result is the news media ranking: how often each newsroom is first in the stories it covers, and how far behind it usually is when it isn't.

4. The importance score

Each story gets a score from 0 to 100. It's the sum of seven signals, each with a ceiling, so no single one can carry a story on its own:

The importance score says how much a story matters. It doesn't say whether it's true. That depends on its sources, which are always linked.

5. Market moves

We follow the 150 largest coins by market value that trade at least $5 million a day. Prices come from CoinGecko every half hour.

A move is flagged when it's at least 3% and at least two and a half times the coin's usual volatility over the last 14 days. We then remove the market: from each coin's change we subtract the median change of the 20 largest coins. What's left is the coin's own move. If that is under 3%, it was the market, not the coin.

For each flagged move we look for what was published before it. Something published before a move is a candidate, not a proven cause. When nothing was published, the page says "No reason found for this move."

6. How we use AI

Language models help in three places:

• Deciding whether two articles describe the same event, when the measurements above aren't conclusive.
• Drafting the short summary of a story.
• Writing the one-line note under a market move. These are marked "model read": they summarise the headlines and are not a verified cause.

Summaries go through checks before they are published:

• Every claim has to cite a piece that really belongs to the story.
• Every figure in a claim has to appear in its source.
• Wording that asserts a cause or reads as investment advice is rejected.
• A second model, different from the one that wrote the draft, reviews it.

A draft that fails is rewritten once with the complaints. About a third of drafts fail the first time.

To draft a summary, the model sees the headlines and short extracts, at most a few hundred characters per article. Nothing about our readers is ever sent to a model.

7. Where the data comes from

8. The limits

• We only see what we read. A newsroom we don't follow can be first without us knowing. The ranking measures our sources, not the whole press.
• Publication times are declared by publishers. We discard the implausible ones, but a wrong time that looks plausible gets through.
• Grouping can be wrong. Two events can be merged, or one split in two. The rules above make it rare, not impossible.
• Prices are taken every half hour. A move that starts and reverses inside that window doesn't show.