The top program summary passed to the LLM can be misleading, because it flatly claims that performance on all metrics was good, even if some of the metrics have bad values that need to be improved upon. Due to the claim that the value is good already, the LLM will be less likely to address the topic.
https://github.com/codelion/openevolve/blob/c779ac9a5d013c1ef8a12fa4ba8869c271b28a0c/openevolve/prompt/sampler.py#L318