Minerva-7B Internal Latent Knowledge Differs From Model Outputs

A white-box auditing framework called LLM endognostics reveals that Minerva-7B maintains robust risk differentiation in its residual stream even when behavioral evaluations show it complying with or refusing both sides of a prompt. Jacobian lens projections show the model internally knows the truth in 63.7% of cases where external behavior fails.