2048 × Jev

It would be, which is why this lab refuses to. Every turn is one Choice over the four directions with no heuristic fallback, so each move stays attributable to the model, and the confidence threshold defaults to 0 so the model plays every move. Quietly swapping in a rule when it is unsure would make it impossible to tell a model move from a scripted one.

No. Jev is a decision model with no /v1/chat/completions, so an OpenAI-compatible text-completion layer does not fit the SystemOne protocol, and fully offline is impossible since the official weights are not released. The README points at a LiteLLM proxy, LangChain's TypeSafeClassifier reading a base URL, or your own /v1/systemone layer, plus Laya as an open-source stand-in that needs a sidecar.

No. A static site has no server functions, so it holds no key and calls the hosted Vercel deployment cross-origin instead, where the key stays in the function process; repoint REMOTE_PROXY_BASE in src/client/api.ts at your own domain. Note the sealing of a local plaintext key is explicitly not cryptographic: the default passphrase lives in src/server/seal.ts.

Yes. Probability distribution, confidence, latency and token cost are surfaced live for every turn, and the log drawer holds each full request and response round trip read-only. The API drawer sets backend, model ID, key and retry behaviour; managed backends hide the endpoint and lock the model ID, while your own backend exposes both.

It runs four steps and aborts if any fails: a DOM id consistency check, then 76 unit tests, then compiles the client and server, then starts the server. The project is pure TypeScript with no bundler, and the server prints local and LAN addresses.

The key is read from VERCEL_AI_GATEWAY_KEY, then a sealed file, then a plaintext file. Sealing with tools/seal-key.ts is explicitly not cryptographic because the default passphrase is in src/server/seal.ts, so it only guards accidental exposure. Real protection comes from keeping the key server-side.

The Strategy context group injects a rule description, a strategy hint and consequence prediction into the request, and the consequence prediction toggle has the biggest effect. The Strategy rules group sets the decision policy and confidence threshold, while the game group changes the board without touching Jev's input.