What is the bot gauntlet?
The bot gauntlet is the set of checks a candidate CL3O version must pass before Hexodic serves it to players. CL3O is Hexodic’s learning bot; the six fixed practice bots are not part of the gauntlet and never change. The gauntlet is the enforcement half of play-to-train: human play sets what a candidate must handle, and the gauntlet keeps promotions honest.
Key terms: a candidate is a proposed set of CL3O weights; the served version is the one players face now; an SPRT (sequential probability ratio test) plays games until it can accept or reject “the candidate is stronger”.
The trigger: a daily run after arena games
A Hexodic learning run is scheduled daily and skips any day on which no arena game against CL3O finished. When it runs, a seeded self-play search makes small changes to the served weights and plays each candidate against the served version. The seed includes the run, so each day explores new candidates while any single run can be repeated exactly.
The gauntlet itself
- Strength test. The best candidate plays an SPRT against the served version, with a capped game budget and one time limit for the whole job. A test cut off by the time limit is discarded, never counted as a pass.
- Exploit replay. Every recorded game a human won against CL3O is replayed to CL3O’s last move. A candidate that fixes fewer of those losses than the served version is rejected, however strong it looked in self-play.
- Promotion. A candidate that clears both checks becomes the next served version, and the arena’s reigning-champion title resets for it.
- Failure. Otherwise the candidate is discarded and the served version stays.
Why so strict?
Because each promotion becomes the baseline for every later candidate, a lucky promotion would carry forward. The SPRT makes the strength claim statistical, and the exploit replay stops a candidate from winning at self-play while forgetting how humans beat it. The same standard applies to the game’s own numbers, whose method and limits are on how Hexodic was measured.
What feeds it, and what doesn’t
The gauntlet uses arena games against CL3O, which are tied to a signed-in account so wins can be verified. The anonymous game records the iOS app can send do not feed the served model. The full data inventory is on how your games help, and national bots — which would be country-trained, and remain vision — are on the roadmap.
Think you can beat the served version? Challenge CL3O or get Hexodic.
Reviewed 2026-09-15.