TL;DR. Clear win WITH the LoRA at 50% — reward 0.844, far above the evolved prompt alone (0.634). This is the strongest LoRA case in the set: at 50% merge nearly every metric improves at once — emo 0.100 to 0.173, genuineness 0.159 to 0.198, blend 0.439 to 0.577, and WER even drops to 0.196. The only real trade-off is the quality proxy, which slips modestly from 0.529 to 0.475. The fatigue read pulls in a low-arousal longing/numbness color (corr Longing 0.62, Emotional Numbness 0.58, corr neg Arousal -0.72); higher merges regress (LoRA150 doubles emo to 0.29 but WER hits 0.666 and reward falls to 0.631).
Without a LoRA
Best: evolved prompt, no LoRA → reward 0.63
GENERAL: A voice intensely expressing fatigue exhaustion, a weary, exhausted voice, heavy and dragging, barely able to stay awake, raw and unfiltered.
SCRIPT: (utterly exhausted, weary and dragging) "<neutral sentence>" · temp 1.1 top_p 0.95 top_k 25
Clear win WITH the LoRA at 50% — reward 0.844, far above the evolved prompt alone (0.634). This is the strongest LoRA case in the set: at 50% merge nearly every metric improves at once — emo 0.100 to 0.173, genuineness 0.159 to 0.198, blend 0.439 to 0.577, and WER even drops to 0.196. The only real trade-off is the quality proxy, which slips modestly from 0.529 to 0.475. The fatigue read pulls in a low-arousal longing/numbness color (corr Longing 0.62, Emotional Numbness 0.58, corr neg Arousal -0.72); higher merges regress (LoRA150 doubles emo to 0.29 but WER hits 0.666 and reward falls to 0.631).