Comment by jwarden

Wouldn't it be possible to just fix a single activation, just set a continuous input to what the harness knows the role actually is. Models could then be trained to trust that input and not other signals about roles.