Open challenges
Standing challenges on Mistboard. Each has a fixed test and a public result. Win an engine challenge and your engine runs on the site as the bot, with your name on the bot page and in the changelog. Win a match and your games go up as a study with your name. Every win gets a post here.
Engines
The bots for these games are open source, hand-written classical engines, and a strong human beats each of them. Beat one in a match and yours becomes the bot.
| Game | Current bot | What wins the seat | Status |
|---|---|---|---|
| Jieqi | PikaJieQi, our build of Pikafish’s jieqi_old branch, no net, 4 s a move (repo, prebuilt binary) |
A net for Pikafish’s jieqi branch, which has a jieqi NNUE architecture and has never had weights to load |
Open |
| Banqi | MistyBanqi at 3.5M nodes a move, Rust, UCI (repo) | Any engine that beats it at the same node budget | Open |
| Jungle | MistyJungle at 5M nodes a move, Rust, UCI (repo) | Any engine that beats it at the same node budget. Flip Jungle on the same terms | Open |
On jieqi, what has been tried: ten nets trained on jieqi_old’s own evaluations (up to 5M positions, depth 8 and 12, one round of self-play), and the best scored 0.4 against its teacher over 200 games. A net that learns the teacher’s evaluation tops out at the teacher; labels from something stronger than jieqi_old are the open path. MistyBanqi has never played the academic Chinese dark chess programs (CLAP_CDC, DarkKnight); a port of one that beats it at the same budget qualifies. MistyJungle beat the strongest open-source Dou Shou Qi engine we could run by about 16 Elo over 200 games, most of them draws.
The match: 200 games against the current bot, 100 as each colour, the same time or node budget per move for both on one machine, seed published. A score of 0.55 or better (about +35 Elo) takes the seat. I run it within two weeks of a submission and post the result either way. Submit by opening an issue on the engine’s repository with a Linux x86-64 binary or a weights file, the exact command line, and your own match log. For jieqi, the trainer, feature extractor and referee we validated against the engine are on the jieqi branch of my nnue-pytorch fork (JIEQI.md there says what was tried).
Fog chess (Misty) and the Fairy-Stockfish xiangqi variants come later.
Beat the bot
| Challenge | Terms | Result | Status |
|---|---|---|---|
| Beat PikaJieQi Strongest at jieqi, best of eight, at a time control you choose | Play signed in on the site so the games are on record. Four wins takes it; a drawn match is a rematch | The eight games go up as a study with your name, and they become the benchmark the jieqi engine work is measured against. People beat it in single games; a match is the benchmark we lack | Open |
Submit by opening an issue on the mistboard repository with links to the games.
Wanted
Someone who plays Hong Kong mahjong for real, to read the scoring table at /rules/mahjong against how their table actually scores. Nobody who grew up playing has checked it, and the variant stays closed until someone has. Named on the rules page. Same issue tracker.