← all emotions · markdown (agent skill)

Disappointment conditioning guide

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

TL;DR. Best reward is without a LoRA, use the evolved steering prompt (BASE_P 0.581, well above BASE 0.489 and every LoRA); LoRA150 is merely the best of the LoRAs (0.478), not the best overall. The LoRA is a genuine intensity dial, target emotion climbs 0.021 to 0.093 and genuineness 0.161 to 0.228, but blend collapses (0.450 to 0.281) and quality craters (0.534 to 0.331), which is why reward never catches the prompt. The key side-effect is that it drags in full-blown Sadness, Distress and Helplessness (all correlating ~0.85), so disappointment tips over into outright grief. Default to the prompt; only reach for LoRA150 when you specifically need audibly stronger, more genuine emotion and can accept the blend/quality loss.

Without a LoRA

Best: evolved prompt, no LoRA → reward 0.58
GENERAL: A voice viscerally expressing disappointment, a deflated, let-down voice, heavy with disappointment, quietly crushed, in every breath. SCRIPT: (deeply disappointed, deflated) "<neutral sentence>" · temp 1.1 top_p 0.9 top_k 40

With the emotion LoRA (dose)

50% — reward 0.46 · emo 0.02
genu 0.16 · blend 0.42 · qual 0.51 · WER 0.43
100% — reward 0.45 · emo 0.07
genu 0.19 · blend 0.36 · qual 0.43 · WER 0.57
150% ★ — reward 0.48 · emo 0.09
genu 0.23 · blend 0.28 · qual 0.33 · WER 0.43

Conditions

conditionrewardemotiongenublendqualityWER
neutral prompt, no LoRA0.490.000.110.450.630.27
evolved prompt, no LoRA0.580.020.160.450.530.20
evolved prompt + LoRA 50%0.460.020.160.420.510.43
evolved prompt + LoRA 100%0.450.070.190.360.430.57
evolved prompt + LoRA 150%0.480.090.230.280.330.43

Side effects & correlations (LoRA @ 100%)

↑ pushed up: Sadness (+0.60), Helplessness (+0.60), Distress (+0.60)
↓ pushed down: Warmth (-0.30), Recording Quality (-0.28), Fullness (-0.28)
+ correlates: Sadness (+0.88), Bitterness (+0.84), Distress (+0.83)
− anti-correlates: Interest (-0.66), Esthetics (-0.59), Structure (-0.53)

Conclusion

Best reward is without a LoRA, use the evolved steering prompt (BASE_P 0.581, well above BASE 0.489 and every LoRA); LoRA150 is merely the best of the LoRAs (0.478), not the best overall. The LoRA is a genuine intensity dial, target emotion climbs 0.021 to 0.093 and genuineness 0.161 to 0.228, but blend collapses (0.450 to 0.281) and quality craters (0.534 to 0.331), which is why reward never catches the prompt. The key side-effect is that it drags in full-blown Sadness, Distress and Helplessness (all correlating ~0.85), so disappointment tips over into outright grief. Default to the prompt; only reach for LoRA150 when you specifically need audibly stronger, more genuine emotion and can accept the blend/quality loss.