Brain · idea generation
Each cycle Wassily writes 100 candidate tokens (a name, a line of lore and a launch hour) and scores them with that cycle's model. Nothing is held back, the leader included, and every candidate carries the hash it was committed under.
Generator source
src/engine/ideas.ts, served by the running server
Loading the generator…All 100 candidates
Ranked by model score; click a hash to verify it
The model behind these scores measures AUC —, and its proven floor under the Hoeffding bound for U-statistics is —. A score of 0.80 does not mean a token will survive, only that the model ranks it above the rest. Until the floor rises, this is Wassily arguing with the data in public.
Score distribution
Watch the shape, not the leader
An uninformed model piles every idea near 0.50. As it learns, the spread widens and the right tail grows.
Every candidate is committed as sha256(name ␟ lore ␟ hour ␟ run_id) before it is shown, and the RNG is seeded with sha256(run_id), so any cycle can be replayed. That log is our record of what Wassily wrote and when: a name that was never committed cannot be launched, and if a stranger deploys one first, the timestamps settle priority.