What is the bot gauntlet?
The gauntlet is hexodic’s promotion system: the tournament a candidate bot must sweep before it becomes the production bot. It’s the enforcement half of play-to-train — the loop where decisive human wins seed new candidates. The seeding is human; the gauntlet is what keeps the results honest.
The trigger: a decisive human win
Candidates aren’t generated on a schedule. The trigger is a player beating the production bot decisively — the system treats that as evidence of an exploitable weakness, locates the pivot ply where the game turned, and generates a candidate variant that accounts for it. No decisive wins, no candidates: the players are the search heuristic.
The gauntlet itself
A candidate doesn’t fight the champion — it fights everyone:
- It plays the entire current agent pool, not just the bot it hopes to replace.
- Against every single pool member it must win with statistical significance, scored by a Wilson 95% confidence gate — enough games that the win rate’s lower confidence bound clears the bar, so a lucky streak can’t sneak a weak candidate through.
- One failure anywhere ends it. A candidate that dominates the champion but drops a matchup to an older bot is discarded — regressions against any known style are disqualifying.
- Clear every bar and the candidate is promoted to production; it becomes the bot players face, and the next thing a decisive win can dethrone.
Why so strict?
Because the loop compounds. Every promotion becomes the baseline every future candidate must beat, so a single generous promotion would poison the pool downstream. The Wilson gate makes strength claims statistical rather than anecdotal, and the whole-pool sweep makes them general rather than style-specific. It’s the same reason the game itself ships with measured numbers — a 0.25% draw rate, 46.38% first-player win rate, branching of 911 / 1,009 / 665 at half-moves 5 / 10 / 20 — the project’s standing rule is that claims get measured or they get labeled.
What feeds it, and what doesn’t
The gauntlet consumes anonymously recorded completed games — moves, outcome, and optionally a region-level country tag. No account, no personal data, no neural network training on you; the mechanism is a tournament of classical search agents. The full data inventory is on how your games help, and where the loop is headed — country-trained bots, clearly labeled vision — is on the roadmap.
Think you can trigger a candidate? The six-tier ladder ends in a deep-searching expert. Get hexodic.