🎮 Gaming
Game agents, NPC behavior, gameplay testing and interactive experiences
Projects in this industry
View all →Game Harness Lab
A local game-agent workbench that tests and evolves Decision harnesses using gameplay feedback, GEPA and paired evaluations.
S1 →StartLux Decision chooses each gameplay action from the supplied action set. The published configuration uses think_mode=never. GLM-5.3 and GEPA improve harness instructions outside gameplay; host code enforces rules, validates choices and executes actions. These are StartLux Decision results, not verified runs of official TypeSafe Jev.
Decision Arena
Play against decision models, compare AI opponents, and share game replays across board and real-time games.
S1 →Jev, StartLux System One, or a compatible decision model selects the next action from candidates supplied by the game engine, using the current game observation and visible context. The engine handles rules, continuous physics, action validation, execution, and replay recording. Unknown or invalid choices stop the turn rather than being replaced with a random move.
Case studies
View all →StartLux vs Jev: One Snake Match
A recorded 20×20 Snake comparison between local StartLux 27B and Jev 1.14.
StartLux + BetterGI: Genshin Task Selection
StartLux selects task-level skills while BetterGI handles routes and combat.
StartLux Roadster Rush Driving Demo
A short demonstration of decision-model control in a browser racing game.