This is the real engine. Unlike a prompt-engineered simulation, every response below is generated by a live language model whose internals are being modified in real time — Contrastive Activation Addition (CAA) steering vectors added to the residual stream, per-head attention scaling, and KV-cache decay. Drag a compound onto the avatar, set the dose, and chat. Intensity is a multiplier: past 1.0 you are overloading the pack — expect it to get strange.