Supplementary data for "Emotion and authorization steering both move cheat; trained-probe suppression doesn't undo it: a mechanistic study in Gemma-2-2B"

This dataset contains the supplementary response, judge-label, activation, and intermediate data for the preprint "Emotion and authorization steering both move cheat; trained-probe suppression doesn't undo it: a mechanistic study in Gemma-2-2B" by Dmytro Tarasenko (2026). Contents: an 880-record emotion-steering sweep across four cheat surfaces in Gemma-2-2B-IT with Sonnet-judge labels and cached layer-17 activations for 850 of those records; a 1800-record T2/T3 cross-surface auth_dir sweep; a 600-record T4 supplement; 390 v3 auth_dir runs (on-target dose-response, F8 benign control, superseded cross-surface T4); and figure-support intermediates (cell-LOCO probe results, Phase A direction-suppression summary, sweep projections). See MANIFEST.md inside the archive for the full per-file record-count and paper-section mapping, and README_DATA.md for usage and citation guidance. Companion code, prompts, and steering vectors are at the GitHub repository linked above.