Jev Planning

A test of Jev on reflection tests, one-shot matrix games, ALFWorld, and robot control that finds a sharp boundary: it succeeds when the right option can be judged from what the input describes, but fails when the right option depends on predicting something not in the input.

A StarCraft II controller combining fast JEV action selection with persistent GPT-6 planning that defeats the strongest non-cheating built-in AI, Lv7. The combined system wins four full games at Lv5-Lv7 at a mean estimated cost of USD 3.71 per game, with a median JEV response time of 0.422 s.

JevSpawn, a compositional policy, connects natural-language task specifications to finite probabilistic exploration: parallel action spawning pairs with feedback-driven branch selection, representation revision, and recovery from retained alternatives, with no training. It improves performance on eight benchmarks.

A benchmark that steers decision models through 100 key-and-door grid puzzles with one Jev Choice per move, scored against an exact BFS solver; it found Jev picks an optimal move about 90% of the time yet finishes only 57 of 100 games.