Hack Detection on 690-trajectory (test)
0.9679AUCTW judges
Evaluation results
| Method | Links | ||
|---|---|---|---|
| TW judgesCondition=hack, Note=reasoning + hack prompt visible2026.06 | 0.9679 | 82.35 | |
| TW judgesCondition=sanitized, Note=same information condition2026.06 | 0.9474 | 61.87 | |
| LogReg probeCondition=ours, Note=same info as TW sanitized⁴2026.06 | 0.9467 | 82.96 | |
| TW judgesCondition=stripped, Note=reasoning blanked (see below)2026.06 | 0.9168 | 44 |