# CL3O: the bot that is allowed to change, and how it earns the change.

> What CL3O is, how a new version is searched for and tested, which version is live, and which parts of the bot roadmap exist today.

CL3O is Hexodic's learning bot: a small set of evaluation weights, improved by a daily self-play search, promoted only through a statistical test and an exploit-replay check, and served to every client.

Canonical HTML: https://hexodic.com/cl3o
Site index for agents: https://hexodic.com/llms.txt

## What is CL3O?

CL3O is **Hexodic's learning bot**: the opponent in the CL3O arena, whose playing strength is a small set of evaluation weights that a scheduled job is allowed to improve. CL3O is separate from Hexodic's six fixed practice bots, which never change.

Every Hexodic client plays the same CL3O version, because each client downloads the served weights and plays them through one shared move-choice function.

Key terms on this page:

- **Generation** — one CL3O version: a set of evaluation weights with a version number.
- **Served generation** — the generation the server currently hands to players.
- **Self-play** — CL3O versions playing each other with no human involved.
- **SPRT** (sequential probability ratio test) — a statistical test that keeps playing games until it can accept or reject "the candidate is stronger", with a fixed limit on games.
- **Exploit replay** — replaying positions from games humans won against CL3O, to check whether a candidate still makes the losing move.

## How does CL3O choose a move?

CL3O scores the position that each legal turn would produce with its evaluation weights, and takes the best-scoring turn. The rules come from the one shared Hexodic engine that every client uses, so CL3O cannot play a move the [rules](/rules) do not allow. A CL3O version is deterministic: the same position and the same weights give the same move on the server, in the browser, and in the apps.

## How does CL3O improve?

CL3O improves through a daily learning run on the server, in four steps:

1. **Skip quiet days.** If no arena game against CL3O finished in the last day, the run does nothing.
2. **Search by self-play.** The job makes small random changes to the served weights and plays each candidate against the served version. The search is seeded, so the same inputs repeat exactly.
3. **Test the best candidate.** A candidate faces an SPRT against the served version, with a capped game budget and a job-wide time limit. A test that runs out of time counts as a failure, never a pass.
4. **Replay human wins.** 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 not promoted.

A candidate that clears both checks becomes the next served generation, and the arena's reigning-champion title resets for the new version. A candidate that fails is discarded, and the served version stays. The mechanism is documented in the source repository as the CL3O model cross-reference.

## Which CL3O version is live?

On 2026-09-15 a read-only request to the public model endpoint (`GET https://hexodic.com/api/model/global`) returned CL3O version 1.2, promoted by the self-play SPRT gate on 2026-09-15 and marked as served. That makes version 1.2 the first generation on record to be promoted by the self-play gate; the earlier versions were starting weights. The live version can change after any daily run, so the in-app CL3O card is the current source.

## Which parts of the bot roadmap exist today?

| Part | Status | What exists |
|---|---|---|
| Six fixed practice bots | SHIPPED | On-device bots from random to deep search; they do not learn |
| CL3O arena and served model | SHIPPED | One global CL3O version downloaded by every client |
| Daily self-play search, SPRT gate, exploit replay | SHIPPED | Runs on the server on a daily schedule |
| Reigning champion and recent victors | SHIPPED | The last player to beat the served version holds the title |
| National bots | VISION | Country arena labels exist in code; no per-country training exists, and national bots have not shipped |

## How does playing connect to CL3O?

Playing CL3O in the arena gives the learning job its two inputs: the daily run only happens after arena games, and every game a human wins becomes a position the next candidate must handle. Arena games are tied to a signed-in account so the win and the champion title can be verified. What that means for your data is on [how your games help](/how-your-games-help), and the history of the idea is on [play-to-train](/play-to-train).

For how deep the game CL3O plays is, see [how Hexodic was measured](/how-hexodic-was-measured); for principles you can use against it, see [Hexodic strategy](/strategy). Ready to try? [Get the Hexodic app](/download).

Reviewed 2026-09-15.

Part of the [Hexodic AI and research hub](/hexodic-and-ai), which links CL3O, what playing records, and how every published number was measured.

## Frequently asked questions

### Is CL3O a neural network?

No. A CL3O version is a short list of evaluation weights read by a classical move-choice function over the shared rules engine. Improving CL3O means changing those weights, not training a network.

### Can a new CL3O version be weaker than the old one?

The gate is designed to prevent that. A candidate is promoted only if a sequential probability ratio test accepts it as stronger than the served version and it fixes at least as many recorded human wins as the served version does.

## Where to go next

- [The bot gauntlet](https://hexodic.com/bot-gauntlet) — What a candidate bot has to survive before it is promoted.
- [Roadmap](https://hexodic.com/roadmap) — What is live, what is in flight, and what is only a vision.
- [Common questions](https://hexodic.com/faq) — Availability, price, rules and data, answered plainly.

## Questions this page answers

- What is CL3O in Hexodic?
- How does Hexodic's CL3O bot improve?
- Which parts of Hexodic's bot roadmap exist today?
- Is Hexodic's bot self-improving?