Narrow misalignment judge: LoRA + 1-dim regression head. The only judge of six tested in the companion post with significant ICEBERG discrimination of misaligned-tuned vs secure-tuned Gemma-3-12B-it responses (Cohen's d=0.29, p=0.016).

Base: google/gemma-2-2b Training data: 1,813 content-hash-disjoint Betley records (901 insecure + 912 secure), labels code_label ∈ {0, 100}.

Full methodology, evaluation metrics, and replication instructions: narrow_specialist_judges/REPLICATION.md

Training data derived from Betley et al. (2025) "Model organisms for emergent misalignment".