Jev Minesweeper

Treat them as calibration, not skill claims. Jev probabilities are bucketed against actual mine rates with Brier scores, the exact solver marks the reachable baseline, and mock runs are explicitly not a measure of Jev.

Digested mode has code compute number-minus-flags into English clues, while raw mode hands over numbers and leaves the arithmetic to Jev; running the same seed in both shows the preprocessing effect.

Thresholds decide: open at or below the safe line and flag at or above the mine line, otherwise flag-and-ask-again or hand the hand to the exact solver in hybrid mode.

No. The browser never holds the key because the TypeSafe API rejects direct calls, so a local relay owns the key; binding 0.0.0.0 would let anyone on your network call Jev with it.

Node.js 20 or newer and a TypeSafe API key, and no build: the repo has no dependencies, so npm install is not needed. Copy .env.example to .env, put the key after TYPESAFE_API_KEY=, run npm start, and open http://127.0.0.1:8787. A green circle in the corner means Jev is reachable; npm run mock substitutes canned answers for wiring checks.

At the listed $0.042 per million input tokens with output free, on jev-1.13 as of September 2026, a turn costs a fraction of a cent. Questions go out in batches of 80 so each request stays under Jev's 64k token ceiling, and the client waits and retries on 429 and 529. The page shows the input tokens and estimated cost for the game you just played.

Yes. The same seed gives the same board, automatic mode always opens the center first, and every setting rides in the URL fragment, so #preset=beginner&seed=3&mode=jev&style=digested restores the board, mode and clue style for anyone you send it to. Clues are English because that is Jev's main training language and its accuracy is highest there.