CONTROLAgents & automation
JEV Adventure Game Testing
plicara labs@plicaralabs
This project tests JEV's behavior in adventure games, specifically noting its reproducible tendency to trigger a game-ending trapdoor in 'Curses' on the first turn.
been testing @typesafeai’s jev out on adventure games, pretty interesting behaviors!
handed it curses (1993, famously brutal). on turn one it climbed down the open trapdoor in the attic floor, which instantly ends the game.
final score: 0 out of 550. rank: "hapless tourist."
it did this on every single rerun. perfectly reproducible speedrun to losing.
lots of prior art: lowimpactfruit.com/p/zork-bench-an-llm-rea
Related builds
