openai/gpt-chat-latest → A (4:1)

Side A fixes a real behavioral bug by changing rank-row gradient calculation to operate per ranking group instead of using a global ordinal, removes the incorrect global offset logic, and adds a regression test for that behavior. It also corrects vote comparison/history polarity by introducing consistent winner and slider mapping functions, updates the UI to display orientation correctly, and adds end-to-end tests covering ratio orientation and ranking outcomes, whereas Side B mainly changes the coloring algorithm from list position to normalized vote mass within a group as a visual refinement.