Contents
Crawlers
Each crawler is a headless browser with a small brain. It takes a URL from its queue, loads the page (or downloads the PDF), pulls out the readable text, and asks the brain three things: is this about crypto, how good is it from 1 to 5, and which links on it are worth reading next.
A page is kept only if it is relevant and scores 3 or better, and only if it is not a copy of anything already kept: exact copies by SHA-256, near-copies by MinHash (Jaccard similarity above 0.8).
Crawlers respect robots.txt, never hit one domain more than once every 2 seconds across the whole fleet, cap their pages per hour, and skip a blocklist of social and price sites.
The queen ladder
The dataset grows toward a ladder of model sizes, v1 to v8. Each rung needs params × 10 ÷ 4 unique tokens: 10 tokens per parameter, at most 4 passes over the data. Crossing a rung starts a training run on its own.
Hives
Every coin launched here gets its own vault as the coin's creator, so its creator fees belong to that hive alone. They are split 50% to the hive's crawling and training, 20% to people who spawn crawlers in it, 20% to buy back and burn $HIVE, and 10% to whoever launched it.
$HIVE
Burn 1,000,000 $HIVE to spawn a crawler you own. Creator fees are claimed daily: 20% to active crawlers, 60% to training, 20% to the queen's trading wallet.
Parts
- hive
- web
- site + WebSocket relay
- launch API
- crawler worker
- browser contexts
- brain (DeepSeek)
- dedupe (SHA-256 + MinHash)
- postgres
- redis
- storage (R2 or volume)
- web