Frame-Conditioned Moral Computation in LLaMA 3.1-8B-Instruct: A Mechanistic Interpretability Audit of Ethical Reasoning

In three linesMechanistic interpretability audit of LLaMA 3.1-8B-Instruct on 54 moral prompts using Transluce platform. Reveals Situational Anchor Effect: domain-specific representations dominate activation rankings regardless of ethical content. Ethics capacity remains constant but salience is highly sensitive to prompt's interpretive frame. Identifies candidate ethics neuron (L16/N3837) stable across temperatures.

Summary generated by Claude — human-verified