Mechanistic interpretability

Eval-Awareness at Deployment Base Rates

A linear probe can separate EvalAwareBench from WildChat with AUROC 0.986 — and still produce 99% false alarms at a 1% …