FSRS's counterintuitive punchline: when is the optimal moment to review?

Right when you're about to forget — the biggest stability gains come after R has already dropped.

If you review while R is still ~100% (you clearly remember), the successful recall barely teaches the model anything and barely boosts stability. Let R fall toward your target first, then recall successfully, and S jumps much more — you buy a longer next interval from the same single review.

This is the whole efficiency argument: slightly uncomfortable recall, scheduled deliberately, is the cheapest durable learning there is. Easy reviews feel good and accomplish little.