MIT SwarmWorld findings replicated on four RTX 3090s with open Qwen model

HAL8's AI Local Lab reproduced the core results of MIT's SwarmWorld experiment using a quantized Qwen3.8-27B model served via vLLM on four RTX 3090 GPUs. The run took about 26 hours and 12,802 model calls; nine of 12 paired effects matched MIT's direction, and portfolio resilience was even stronger.

• Nine of 12 paired effects matched MIT's original direction
• Solo model hit 6 of 16 milestones; in full society it hit all 16
• Two 4-bit builds of the same model scored 0.41 vs 0.24 in validity pilot
• Six of seven released MIT traces failed a hidden-state check