Support for pairwise comparison #128
Description
There are domains where it is easier to compare head-on 2 instances of the evolution candidate, instead of providing a scalar score. Due to GEPA's focus on pareto frontiers, which can be found by simple direct comparison between elements, GEPA should be able to accept a scoring function, that provides pairwise comparison between candidates.
This could potentially be addressed by the Bradley-Terry model.
There can also be experiments that convert a scalar metric to be a pairwise score, and compare the 2 techniques.
Context: https://discordapp.com/channels/1161519468141355160/1410016143782973652/1432352069708152955