← all emotions · markdown (agent skill)

Confusion conditioning guide

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

TL;DR. Best reward comes with the LoRA at 50% (0.521), beating BASE_P (0.496) and BASE (0.490), while higher merges fall apart (LoRA100 0.344, LoRA150 0.372). Recommendation: use the evolved prompt plus a 50% LoRA. The gain is a genuineness lift (0.140 to 0.172) with blend held roughly steady (~0.423); the trade-off is a small quality drop (0.551 to 0.503) and higher WER (0.196 to 0.302). One caution: the evolved prompt alone tends to make confusion sound irritated rather than doubtful (BASE_P correlates positively with Anger/Teasing and negatively with Doubt at -0.51), whereas the LoRA instead pulls in Fatigue and Fear and suppresses Concentration.

Without a LoRA

Best: evolved prompt, no LoRA → reward 0.50
GENERAL: A voice powerfully expressing confusion, a baffled, disoriented voice, struggling to make sense of things, hesitant and lost, pouring out. SCRIPT: (confused, bewildered, thrown off) "<neutral sentence>" · temp 1.0 top_p 0.9 top_k 30

With the emotion LoRA (dose)

50% ★ — reward 0.52 · emo 0.00
genu 0.17 · blend 0.42 · qual 0.50 · WER 0.30
100% — reward 0.34 · emo 0.00
genu 0.13 · blend 0.31 · qual 0.47 · WER 0.33
150% — reward 0.37 · emo 0.03
genu 0.18 · blend 0.30 · qual 0.41 · WER 0.50

Conditions

conditionrewardemotiongenublendqualityWER
neutral prompt, no LoRA0.490.000.110.450.630.27
evolved prompt, no LoRA0.500.000.140.430.550.20
evolved prompt + LoRA 50%0.520.000.170.420.500.30
evolved prompt + LoRA 100%0.340.000.130.310.470.33
evolved prompt + LoRA 150%0.370.030.180.300.410.50

Side effects & correlations (LoRA @ 100%)

↑ pushed up: Impatience and Irritability (+0.38), Embarrassment (+0.28), Fear (+0.25)
↓ pushed down: Concentration (-0.28), Fullness (-0.19), Structure (-0.18)
+ correlates: Shame (+0.39), Contemplation (+0.36), Doubt (+0.35)
− anti-correlates: Harmonicity (-0.49), Articulation Clarity (-0.41), Attack (-0.35)

Conclusion

Best reward comes with the LoRA at 50% (0.521), beating BASE_P (0.496) and BASE (0.490), while higher merges fall apart (LoRA100 0.344, LoRA150 0.372). Recommendation: use the evolved prompt plus a 50% LoRA. The gain is a genuineness lift (0.140 to 0.172) with blend held roughly steady (~0.423); the trade-off is a small quality drop (0.551 to 0.503) and higher WER (0.196 to 0.302). One caution: the evolved prompt alone tends to make confusion sound irritated rather than doubtful (BASE_P correlates positively with Anger/Teasing and negatively with Doubt at -0.51), whereas the LoRA instead pulls in Fatigue and Fear and suppresses Concentration.