The authors point out in this paper that the literature on Spaced repetition memory system has mostly focused on finding the schedule regime which produces the highest recall rates, irrespective of time costs. The Testing effect is powerful; more practice will produce more stability. So how should one think about these trade-offs?
It’s not even just a matter of needing to review more often: they note that lower retrievability rates correlate with longer response times; and it takes extra time to read the correct answers when you’re wrong.
They run a scheduler with simulated students (?!) who have a fixed amount of time to study, optimizing a threshold for a scheduler which presents items which have fallen below some estimated recall probability level. They find an optimum of 94% for English–Japanese flashcards, then replicate that result with live students.
They find a d=0.64 effect for this threshold scheduling relative to uniform spacing.
A funny thing about this is that my smooth-brain approach to SRS scheduling was already to use a 90% threshold.