# A strategy game whose bots get stronger from real human play.

> Beat the bot decisively and your game is analyzed to build a candidate that has to beat every current bot — by a strict statistical test — before it ships.

Play-to-train means the act of playing improves the opponent: hexodic analyzes decisive human wins, spawns candidate bots, and promotes only candidates that beat every current bot by a strict statistical test.

Canonical HTML: https://hexodic.com/play-to-train
Site index for agents: https://hexodic.com/llms.txt

## What is play-to-train?

**Play-to-train** is a game design in which ordinary play makes the game's opponent stronger: the system studies how humans beat its current bot, generates a candidate that accounts for the exploit, and promotes it only after it proves itself. In hexodic — the worked example of the idea — that proof is statistical, transparent, and adversarial, not a black-box training run.

Deeper on this cluster: [games where playing trains the AI](/games-that-train-ai) — the lineage from citizen science to hexodic ·
[the bot gauntlet](/bot-gauntlet) — the exact promotion mechanics ·
[what's collected, and what isn't](/how-your-games-help).

## Is there a game where playing it trains the AI?

Yes — hexodic, and it's worth being precise about what that means, because the mechanism is *not* a neural network hoovering up your data. hexodic's bot-improvement loop is a **classical search-agent evolution system**: transparent, statistical, and adversarial.

## How the bot actually improves

The loop, precisely:

1. **You beat the production bot decisively.** A decisive human win is the trigger — the system treats it as evidence the bot has an exploitable weakness.
2. **The game is analyzed for the pivot.** The analysis looks for the ply where the game turned — the decision the bot got wrong.
3. **A candidate agent is spawned.** A new bot variant is generated to account for that weakness.
4. **The candidate must conquer the pool.** It plays the *entire* current agent pool, and it must beat every current bot with statistical significance — a Wilson 95% confidence gate, not a vibe check.
5. **Only then is it promoted.** A candidate that clears the gate becomes the production bot. One that doesn't is discarded.

No deep learning, no black box, no training run on a GPU farm — a tournament of programs, seeded by the specific ways humans out-think the current champion. If you can reliably beat the top bot, you are quite literally the selection pressure. The full promotion mechanics — triggers, the gauntlet, the statistical gate — are written up on [the bot gauntlet](/bot-gauntlet).

## What we collect, and what we don't

Completed games are recorded **anonymously** — no account, no personal data. The record is the game itself: the moves, the outcome, and optionally a region-level country tag (never precise location). That's what feeds the loop above. The full inventory, the retention picture, and where the authoritative privacy policy lives are all on [how your games help](/how-your-games-help) — we'd rather over-explain this than have you wonder.

## Where it's headed: national bots

The vision — clearly labeled as vision, because it hasn't shipped — is **bots that represent and play *for* their countries, trained by their own player bases**. Imagine a bot that plays the way a nation's players collectively taught it to, competing on their behalf. Today, your games train one shared bot; the country tag exists so that, someday, they won't have to. Follow the [roadmap](/roadmap) for where that stands.

And if what drew you here is the game itself rather than the loop — hexodic is a deterministic, zero-luck [abstract strategy game](/abstract-strategy-game) first. The bots are the twist, not the point. [Get hexodic](/#get-hexodic).

## Questions this page answers

- What is play-to-train?
- Is there a game where playing it trains the AI?
- How does hexodic's bot learn from players?
- What game uses player matches to improve its bot?