https://www.tidymodels.org/start/recipes/#predict-workflow
In roc_curve, make a mention of the event level being default the first.
I have run into more than once using a factor truth with levels "FALSE" and "TRUE", roc_curve(truth, .pred_FALSE) or roc_curve(truth .pred_TRUE, event_level="second") should be used. This might be backwards to how people may expect roc_curve(truth, .pred_TRUE) to use the "TRUE" level.