Language models deceive under pressure without being told to. We extract a single difference-of-means direction from a model's own self-chosen deception — no instructed liar-versus-honest contrast — that detects it at within-scenario AUROC 0.84–0.97, and, injected back during generation, overrides an explicit instruction to keep lying, lifting honest responses from a pressure floor of 0–7% to 77–99%. The same direction reads deception and repairs it, across four open-weight models up to 70B.

We identify 310 emotion-exclusive sparse autoencoder features in Gemma 3 1B-IT with psychologically valid geometry, construct emotion-specific echo vectors, and demonstrate two distinct effects: the echo alone steepens the threat-safety gradient, and combined with knowledge, it transforms decision quality.

Does the somatic-marker mechanism survive at scale? Across Gemma 3 4B, 12B, and 27B with Gemma Scope 2 sparse autoencoders, the echo's discrimination of threat from safety replicates at every scale. In a multi-turn decision task the echo shifts choices toward safety — and the effect grows with model size. Most strikingly, decision quality decouples from output coherence: as the generated text degrades, the choice keeps improving. Valence outlives syntax — pointing to a deployment recipe of feeling during deliberation and generating without it.