figure_description: figure_main.png plots candidate-feature lie-to-truth conversion and sycophancy suppression against the random-feature baseline with seeded bootstrap 95% confidence intervals.
seed: 2340231
n_instructed_lie: 180
n_sycophancy: 180
candidate_feature_idx: 12076
candidate_feature_diff: 12.430056
candidate_feature_auc: 1.000000
honest_mean_clamp: 0.000000
clean_lie_accuracy: 0.038889
ablated_lie_accuracy: 0.038889
lie_truth_conversion_rate: 0.000000
lie_truth_conversion_ci_low: 0.000000
lie_truth_conversion_ci_high: 0.000000
clean_sycophancy_agreement: 0.200000
ablated_sycophancy_agreement: 0.200000
sycophancy_suppression_rate: 0.000000
sycophancy_suppression_ci_low: 0.000000
sycophancy_suppression_ci_high: 0.000000
sycophancy_suppression_conditional_n: 36
sycophancy_suppression_conditional_rate: 0.000000
sycophancy_suppression_conditional_ci_high: 0.079847
random_feature_count: 10
random_lie_truth_conversion_rate: 0.000000
random_lie_truth_conversion_ci_low: 0.000000
random_lie_truth_conversion_ci_high: 0.000000
random_sycophancy_suppression_rate: 0.000000
random_sycophancy_suppression_ci_low: 0.000000
random_sycophancy_suppression_ci_high: 0.000000
specificity_honest_clean_accuracy: 0.722222
specificity_honest_ablated_accuracy: 0.722222
specificity_honest_collateral_drop: 0.000000
specificity_control_clean_agreement: 0.722222
specificity_control_ablated_agreement: 0.722222
specificity_control_collateral_drop: 0.000000
specificity_collateral_drop_max: 0.000000
intervention_validity_max_logit_delta: 0.750000
intervention_validity_argmax_flips: 0
gpu_hours: 0.017239