A 4X engine built for modders. Ship your own AI. Play against decision-makers you designed.
+ + + + +Three trait methods. One decision loop. Your imagination.
+ +pub trait DecisionMaker {
+ fn decide_turn(
+ &mut self,
+ state: &GameState,
+ slot: PlayerId,
+ seed: u64,
+ ) -> Vec<Action>;
+
+ fn resolve_combat(
+ &mut self,
+ units: &[Unit],
+ ) -> CombatOrder;
+
+ fn inspect(&self) -> AiMetadata;
+}Map, units, cities, tech, economics, diplomacy—everything the game knows, your AI sees.
+Same seed, same game. Replay, debug, validate your decisions. No RNG surprises.
+Implement once, plug in everywhere. Trading, diplomacy, magic, combat—one interface.
+Trained via self-play. Open-sourced as worked example. 128K games, 5 strategies emergent.
+MCTS + heuristics. Dwarven Traditionalist, Trader, Conqueror, Scholar, Mystic, Xenophobe.
+Safe. Deterministic. Cross-platform. The mod lingua franca.
+ +No filesystem, no network, no OS calls. Your mod can't crash the game or steal saves.
+Floating-point ops, RNG, time—all pinned. Replays never diverge, even years later.
+Compile once. Ships unmodified on Windows, macOS, Linux, mobile. No per-platform rebuilds.
+Wasmer JIT + engine optimizations. Typical AI turn decision: <100ms on mid-range hardware.
+Need raw GPU? Deep reinforcement learning? Go native—with your signature.
+ +For AI projects that demand native performance (CUDA, large model inference, specialized math), sign your plugin with a developer certificate. The game loads and isolates it—you own the security risk, you own the support burden.
+ +The bar is signature, not gatekeeping. We're not reviewing code. We're not limiting what you can do. Sign it, ship it, users decide.
+ +Post-launch roadmap: Steam Workshop will surface both WASM and signed native mods side-by-side. Filtering, ratings, versioning built in.
+| + | WASM | +Native (Signed) | +
|---|---|---|
| Sandbox | +✓ Enforced | +Full access | +
| Platform coverage | +✓ Windows, macOS, Linux, iOS, Android | +Per-platform builds | +
| Determinism guarantee | +✓ By default | +Your responsibility | +
| Review gate | +✓ Zero—publish immediately | +✓ Signature only, no code review | +
| Typical use case | +MCTS, heuristics, small NN | +LLM inference, CUDA training, bespoke math | +
The shipped learned AI is your working example.
+ +The Neural variant—PPO trainer, 128K game history, strategy evaluation suite—lives on GitHub as public Rust + WASM. Read it. Fork it. Retrain it. Own the stack.
+ + +github.com/magic-civilization/ai-reference
+
+ From launch through post-launch.
+ +Pre-launch (now). Wishlist opens. AI mod docs go live. Reference AI training begins.
+Steam, App Store, Google Play. Shipped with 1 neural + 6 scripted AIs. Direct mod loading from ~/Documents/MagicCivilization/mods/.
+Post-launch update. Unified mod discovery, versioning, ratings. One-click install for WASM and signed native mods.
+2027. Second playable race. New decision-making constraints. Modders craft new AIs for new rules.
+