do a principle sae manipulation? which layers and which token positions? increases with more layers and all token positions? same for jacobian lens compare effectiveness the manipulation that gives 95% six
