Playable worlds for observable AI decisions.
JevTown is a collection of interactive game experiments built around typesafe/jev-1.13. It turns a model’s strategic choices into legible, animated games—without handing the model control of the rules.
The Jev model chooses from engine-generated legal plans. Deterministic TypeScript owns hidden information, routes, movement, clashes, scoring, and winners.
Most AI demos hide the decision behind a response. JevTown puts the decision on the board.
Game state → legal strategic plans → Jev chooses one → engine executes it → spectators see why
This creates an AI experience that is both expressive and inspectable: the model supplies strategy; the system supplies truth.
| World | The challenge |
|---|---|
| Jev Hunt | Two explorers discover a town, race for three hidden treasures, and resolve contested squares with rock-paper-scissors. |
| Jev Run | Two Jevs navigate a tactical maze, competing for a key and gateway while their routes remain engine-validated. |
| Jev Forest | Two Jevs explore a hidden forest, scout terrain, collect relics, and intercept a rival through authored scene landmarks. |
Each world has its own rules and visual language, but uses the same core boundary: no model-generated coordinates, no invented objectives, and no model-decided winner.
- Bounded AI agency — every decision is selected from an engine-generated, legal choice set.
- Deterministic resolution — pathfinding, world discovery, rock-paper-scissors clashes, scoring, and victory conditions remain reproducible application logic.
- Inspectable play — decision feeds expose tactics, confidence, routes, events, and outcomes instead of presenting a black-box simulation.
- Designed interaction — Pixi-rendered worlds, responsive UI, day/night modes, generated ambient sound, and lightweight event effects make the system feel like a game rather than a dashboard.
- Real-time foundation — Firebase Realtime Database projections support public live state; Upstash supports atomic Crowd Pick state and recent replay persistence.
Typesafe Jev
│
▼
legal plans ← decision API → validated tactic
│ │
└──── deterministic game engine ┘
│
live projection / crowd state / replay events
│
spectators
The app is structured around typed contracts, client services, thin validated routes, and adjacent server operations. Zod validates request boundaries; server-only configuration protects provider, Firebase Admin, and Upstash credentials.
Next.js · React · TypeScript · PixiJS · Zod · Firebase Realtime Database · Upstash Redis · Tone.js · ZzFX · Vitest
JevTown is a playable local and preview system. Its live-state, crowd-pick, and replay foundations are in place.
The next production milestone is an authoritative runner: a protected server process that starts matches, invokes Jev, advances the engines, and publishes the shared global state. Until that lands, this repository does not claim continuously running global games.
The application lives in Frontend.
cd Frontend
npm install
cp .env.example .env.local
npm run devOPENROUTER_API_KEY is required for Jev-powered decision previews. Firebase and Upstash credentials enable the optional live-state, Crowd Pick, and replay infrastructure. Keep all non-NEXT_PUBLIC_ values server-side.
JevTown uses typesafe/jev-1.13 through OpenRouter. Create your own OpenRouter API key with access to Typesafe Jev at openrouter.ai/typesafe, then add it locally:
# Frontend/.env.local
OPENROUTER_API_KEY=your_openrouter_keyNever commit this key or expose it through a NEXT_PUBLIC_ variable.
npm run lint
npm run typecheck
npm test
npm run buildJevTown is a product and systems-design experiment by BashBash Labs: making AI behavior useful not by hiding complexity, but by giving people a way to see it, understand it, and play with it.