Chess

Plays chess: Jev against a cheap OpenAI model (gpt-4.1-nano by default). Python does the tactics bookkeeping before any model is asked: every legal move is tagged (CHECKMATE, SAFE capture, TRADEs, THREATENS, PASSED PAWN, BOXES IN their king, BLUNDER, repeats) and sorted so forcing, game-ending moves come first. Both players are told to win fast: trade down when ahead, never shuffle, and break a stalled game. Jev answers one Choice over those tagged moves, plus a Score (who stands better) and a Noul (is a piece hanging). The LLM gets the same tags as a prompt and replies in JSON. Run / Dataset: twelve labelled tactics positions. A variant scores when it plays one of the accepted moves. Play: a full game, move by move, with cost, tokens and the raw request/response of every ply. Measured: accuracy per tactic, latency and cost per move. The no-model heuristic shows how much of the work the tags already do. One back-rank row is a mate-in-2 trap that the 1-ply scan can't see.