Fig 1Seven models, one labyrinthturn 0 · watching Sol
1 GPT-6.1 Sol2 Claude Sonnet 5.53 Gemini 3.7 Flash4 Qwen 3.8 Omni Flash5 DeepSeek V4.1 Flash6 Claude Haiku 5.57 GPT-6 LunaThe watched model's way, in its colourIts scent, which the thing hunts by: the stronger the scent, the stronger the dotsEach run's own thing, dark, in its pawn's head shape and ringed in its colour: it hunts that run alone, and goes when the run endsA pit, each announced by a draftThe hourglass: the 45 turns every run has, running outEach lamp's oil, empty to full, under its model's head: dimmed once its run is overSeed 82 at Stirring, the benchmark's baseline condition: seven models in the same labyrinth, with the same tells and the same 45 turns. Every move is the model's own, replayed by the game's engine from its run record. GPT-6.1 Sol and Claude Sonnet 5.5 came out with the Heart on turn 18, Gemini 3.7 Flash on 24. GPT-6 Luna (21) and DeepSeek V4.1 Flash (33) were caught, each with the stench in the room. Qwen 3.8 Omni Flash walked into the pit its own Focus had just placed to the west (23). Claude Haiku 5.5 was holding the Heart, four rooms from the way out, when its turns ran out. Each run has its own thing, dark and shaped like its pawn's head: it moves by that run's scent alone and leaves when the run ends. The hourglass is the clock every run shares; the gauges are each lamp's oil. The dots are the engine's scent field after each turn for the run being followed, strongest where the trail is freshest or loudest.