same. i happen to be doing data sets for prompt injection. iteratively altering the prompt makes a big,

same. i happen to be doing data sets for prompt injection. iteratively altering the prompt makes a big, big difference in false positives. i subjectively noticed that much longer prompts really degraded performance. i don't like analogies, but i think it's like predicting "likelihood of more than half of the quarters coming up heads" for n=3,7,19, 71, 139, etc... larger inputs to a probabilistic process narrow the outcomes.