If you are doing a ML approach of some sort in conjunction with your SBG efforts, please don't forget to design your units/modules/etc. with explicit time and plans for correctives and enrichment. If you don't do that yuo are just retesting and the gains associated with ML dissapear. Also don't make the retest harder. That's like having a person who can't hit a 60 mph pitch, so you send them to do more practice and then throw 80 mph at them. Jerk store.
You want there to be little relationship between your test and retest opportunities, because you want to address misconceptions and mistakes so that the first performance doesn not accurately predict the subsequent performance (if you have a high test/retest correlation it probably means that means your correctives were ineffective).