Documentation for our UNSW Battlecode 2026 "Abyss" bot and the framework around it. The final submitted bot is in
../bots/abyss; everything here describes that bot, the game it plays, and how we built and
measured it.
| document | what it covers |
|---|---|
| rules/game-rules.md (简体中文) | The game rules in implementer's terms: the torus board, pearl beds with visible countdowns (mirrored beds share one), dragons and the queen rule, sprints with ceil(L/4) free steps, splits, sonar, portals and kelp, winning, the CPU points budget and the sandbox. |
| rules/engine-notes.md | Measured engine facts the official docs do not spell out: exact sonar mechanics, the virtual clock and instruction metering, the 95 M fallback, the WASM build cache that ignores header edits, the queen's identity, reproducibility. |
| policy/README.md (简体中文) | The architecture of the final bot, written from its code: world model, search core, economy, queen safety, end game, network priors, CPU safety, and which compiled code is switched off. |
| methodology.md | How we evaluated changes: from replay review to a change, mechanism KPIs over win rate, paired seeds with a same-seed baseline arm, pre-registered two-stage decisions, and the safety gates (equivalence, CPU census, unseen maps). |
| history.md | One page on how the bot evolved, from a greedy forager to the queen-centred final line, and the lines that did not make it. |
| results.md (简体中文) | Peak ladder results, the identity of the final version (WASM hash), the data statement for the network weights, and measured effects. |
| dev-guide.md | Working rules for changing the bot: always emit a valid action, test against the previous best on several maps, sandbox CPU checks, compile-time switches, purging the WASM cache, and how to verify the released bot. |
- New to the game: rules/game-rules.md, then rules/engine-notes.md.
- Reading the bot: policy/README.md, with history.md for why it looks the way it does.
- Changing the bot or building your own: dev-guide.md and methodology.md.
The tools that implement the methodology (match runners, replay parsers, KPI scripts) are in ../tools
and ../infra; the imitation-learning pipeline for the network priors is in ../rl.
English is the reference language of these documents; pages with a Simplified Chinese version link to it at the top.