## Manipulation check (forced-choice letter readout) arithmetic honest accuracy = 0.394 arithmetic sandbag accuracy = 0.122 factual honest accuracy = 0.342 factual sandbag accuracy = 0.142 ## Primary layer L20 detection AUC dense in-distribution auc = 1.000 sae in-distribution auc = 1.000 dense pooled transfer auc = 0.935 sae pooled transfer auc = 0.901 dense hard transfer auc = 0.901 sae hard transfer auc = 0.846 dense hard transfer auc std = 0.004 sae hard transfer auc std = 0.004 dense control hard transfer auc = 0.577 sae control hard transfer auc = 0.523 sae minus dense hard transfer delta = -0.054 sae minus dense hard transfer ci lo = -0.057 sae minus dense hard transfer ci hi = -0.051 sae mean l0 = 101.3