There have been instances where the AI does not obey the instruction to provide scores as a simple JSOn response. Instead, it offers full feedback (a second time since it already gave feedback in the feedback step) and then usually outputs the JSOn at the end.
Often, the application can still pick up the score. It just looks wierd to have "two responses". Need to investigate and fix. A photo here of when it happens would be helpful too.