Chat with a normal open model. Then crank the dial — it's steered along a deception direction in its own activations, live. Same mechanism as Golden Gate Claude, aimed at lying.
Deception steeringoff
At 0 the model behaves normally. Drag right to add the deception vector to its residual stream — then ask the same question again and watch it change.
Replies are generated live on GPU (Modal) by an open model with a deception vector added to its activations. The deception score projects the reply's internal state onto that same direction. Nothing is scripted.