What is play-to-train?
Play-to-train is a game design in which ordinary play shapes how the game’s opponent improves. In Hexodic, the learning opponent is CL3O: its evaluation weights are searched for daily by self-play, and a new version is promoted only when a statistical test accepts it as stronger and it handles the games humans have won against it.
Hexodic’s play-to-train loop is classical search and statistics, not a neural network.
Deeper on this cluster: CL3O, the learning bot — the full mechanism and which version is live · the bot gauntlet — the promotion checks · games where playing trains the AI · what is collected.
Is there a game where playing it trains the AI?
Yes — Hexodic, with a precise meaning. Human games do not become training data for a network. Hexodic’s arena games against CL3O decide whether the daily learning run happens and what a candidate must survive: every recorded human win is replayed, and a candidate that fixes fewer of those losses than the served version is not promoted.
How CL3O actually improves
- Players challenge CL3O in the arena. If no arena game finished in the last day, the daily run skips.
- A self-play search proposes weights. Small seeded changes to the served weights are played against the served version.
- A sequential probability ratio test (SPRT) decides. The best candidate must be accepted as stronger than the served version within a capped number of games and a time limit.
- Human wins are replayed. The candidate must fix at least as many recorded human wins as the served version.
- Only then is it served. Every client downloads the new version; otherwise the served version stays.
On 2026-09-15 the public model endpoint reported CL3O version 1.2, promoted by the self-play gate, as the served version. The details and the probe are on CL3O.
What we collect, and what we don’t
Two separate things happen. Arena games against CL3O are tied to your signed-in account, so a win and the champion title can be verified. Separately, the iOS app can send anonymous completed-game records that are never linked to an account and can be switched off in Settings. The full inventory is on how your games help.
Where it’s headed: national bots
The vision — clearly labeled as vision, because it hasn’t shipped — is bots that would represent and play for their countries, trained by their own player bases. Today there is one global CL3O, and no per-country training exists. Follow the roadmap for where that stands.
And if what drew you here is the game itself — Hexodic is a deterministic, zero-luck abstract strategy game first, with strategy principles that follow from its rules. Get Hexodic.
Part of the Hexodic AI and research hub, which links CL3O, what playing records, and how every published number was measured.