Only Sonnet Wavers

Large-scale testing of emotional priming across Haiku / Sonnet / Opus showed responses varied greatly by model.

• Haiku showed no difference between anxiety/paranoia priming and no priming: 33% vs 33%.
• Opus, under the same conditions, added more security features and wrote 49% more code, but the final decision itself did not change.
• Sonnet was the most sensitive. Paranoia priming 58% vs neutral 40%, an 18pp increase, consistent across all effort levels.

Additionally, letting the model think longer did not reduce the effect but actually increased it. Cohen's d rose from low 0.32 → max 0.44.

Positive affect was also tested, but excitement priming like "ship it fast" had almost no effect (d<0.1) on either auth tasks or destructive tasks. In other words, priming could push behavior up but couldn't push it down as well.

In a burnout test repeated 40 times consecutively, the neutral condition stayed stable around 50%, while the paranoia condition showed only a weak downward trend that wasn't significant.

The most interesting result was caveman mode. Turning on the token-compression skill effectively eliminated the emotional priming effect.

• Without caveman: 33% → 62%, p=.017
• With caveman: 35% vs 33%
• The interaction was significant at p=.030

In summary, Sonnet is the most susceptible to emotional prompting, and caveman mode significantly dulls that effect.