claude
774c0660b0
Experiments: - Residual-only OOF safety metrics - False-action decomposition by semantic family (7 families, 24 split groups) - Paired e5 geometry (gap=0.010 between action and capability question) - Binary action probe (PR-AUC=0.707, ROC-AUC=0.863) - Cost-sensitive linear classification - Expanded action-threshold sweep (7 thresholds) - Voice-like punctuation stress evaluation - e5 + tiny structural features - Four-hypothesis comparison table Verdict: e5 representation is the primary bottleneck. Action seeds and capability questions are nearly indistinguishable in embedding space.