TL;DR. Best with the LoRA at 50% (reward 0.585), but note the plain neutral prompt (BASE, 0.494) beats the evolved steering prompt (0.470) — the steering prompt hurts here, so use BASE + 50% merge. Sourness never registers numerically (emo ~0 in every condition), so LoRA50's win comes entirely from higher blend (0.485) and lower WER (0.179), not added emotion. The price is quality: the proxy drops from 0.625 (BASE) to 0.489, and delivery leans on Disgust/Contempt/Embarrassment correlations, so let the text carry the sour meaning and use the LoRA only for delivery; avoid 100%+ (reward collapses to 0.32).
Without a LoRA
Best: neutral prompt, no LoRA → reward 0.49
GENERAL: A voice intensely expressing sourness, a sour, grumbling voice, peevish and put-out, curdled with displeasure, with every word.
SCRIPT: (sourly, grumbling and peevish) "<neutral sentence>" · temp 1.2 top_p 0.9 top_k 30
Best with the LoRA at 50% (reward 0.585), but note the plain neutral prompt (BASE, 0.494) beats the evolved steering prompt (0.470) — the steering prompt hurts here, so use BASE + 50% merge. Sourness never registers numerically (emo ~0 in every condition), so LoRA50's win comes entirely from higher blend (0.485) and lower WER (0.179), not added emotion. The price is quality: the proxy drops from 0.625 (BASE) to 0.489, and delivery leans on Disgust/Contempt/Embarrassment correlations, so let the text carry the sour meaning and use the LoRA only for delivery; avoid 100%+ (reward collapses to 0.32).