← all emotions · markdown (agent skill)

Astonishment Surprise conditioning guide

Target: EmoNet Astonishment_Surprise strength. reward = (genu + blend + 1.25·emo)/(1+WER).

TL;DR. Best reward comes without the LoRA — the evolved prompt (BASE_P, 0.496) edges out BASE (0.491) and all LoRA merges (best LoRA is 150% at 0.475). Recommendation: rely on the evolved prompt, which already injects surprise cues (Hope/Enthusiasm +0.36, Interest +0.21); reach for LoRA150 only when you need a stronger emo signal (0.033 to 0.145). The trade-off at 150% is a large intelligibility and quality hit (WER 0.296 to 0.621, quality 0.558 to 0.376). Its key side-effect is that the LoRA leaks surprise into Impatience/Irritability (+1.09) and Anger/Disgust rather than delight, and it suppresses Concentration (-0.66).

Without a LoRA

Best: evolved prompt, no LoRA → reward 0.50
GENERAL: A voice unmistakably expressing astonishment surprise, a voice struck with astonishment, gasping, utterly surprised, eyes wide, pouring out. SCRIPT: (gasping, astonished, taken completely by surprise) "<neutral sentence>" · temp 1.0 top_p 0.95 top_k 25

With the emotion LoRA (dose)

50% — reward 0.45 · emo 0.02
genu 0.14 · blend 0.38 · qual 0.53 · WER 0.32
100% — reward 0.39 · emo 0.03
genu 0.14 · blend 0.31 · qual 0.48 · WER 0.39
150% ★ — reward 0.47 · emo 0.14
genu 0.16 · blend 0.39 · qual 0.38 · WER 0.62

Conditions

conditionrewardemotiongenublendqualityWER
neutral prompt, no LoRA0.490.000.110.450.630.27
evolved prompt, no LoRA0.500.030.150.420.560.30
evolved prompt + LoRA 50%0.450.020.140.380.530.32
evolved prompt + LoRA 100%0.390.030.140.310.480.39
evolved prompt + LoRA 150%0.470.140.160.390.380.62

Side effects & correlations (LoRA @ 100%)

↑ pushed up: Impatience and Irritability (+1.09), Anger (+0.43), Dynamic Arc (+0.26)
↓ pushed down: Concentration (-0.40), Warmth (-0.30), Fullness (-0.21)
+ correlates: Elation (+0.78), Disgust (+0.78), Amusement (+0.77)
− anti-correlates: Smoothness (-0.55), Monologue Style (-0.54), Formal Style (-0.51)

Conclusion

Best reward comes without the LoRA — the evolved prompt (BASE_P, 0.496) edges out BASE (0.491) and all LoRA merges (best LoRA is 150% at 0.475). Recommendation: rely on the evolved prompt, which already injects surprise cues (Hope/Enthusiasm +0.36, Interest +0.21); reach for LoRA150 only when you need a stronger emo signal (0.033 to 0.145). The trade-off at 150% is a large intelligibility and quality hit (WER 0.296 to 0.621, quality 0.558 to 0.376). Its key side-effect is that the LoRA leaks surprise into Impatience/Irritability (+1.09) and Anger/Disgust rather than delight, and it suppresses Concentration (-0.66).