SM-2 is a scheduling algorithm. A working recall product is a scheduling algorithm plus a card-generation system plus a review UX plus a lot of edge cases. The algorithm is 5% of the work.
Card generation is the hardest part. Cards must be terse enough to review in seconds but rich enough to test real understanding. Cards must vary in surface so the student cannot recognize them from repetition but hold constant in concept so they test the same skill.
We generate cards by rephrasing problems the student got wrong (or explained back poorly), using Haiku with a strict JSON output. The rephrase preserves difficulty and concept but changes numbers, framing, and (where possible) domain.
Review UX matters more than algorithm accuracy. We flip cards on tap or spacebar. We require a confidence rating before revealing the answer — that data feeds calibration. Four grade buttons (Again, Hard, Good, Easy) mapped to quality 1, 3, 4, 5. Any more granularity and users froze.
Edge cases we hit early: cards that repeat the same surface even after rephrase, cards with ambiguous answer keys, cards that grew stale when a subject was dropped, cards from a session the user no longer remembers. Each got its own small fix — a de-dupe pass, a clarify-answer prompt, a purge-old option, a context banner.
The lesson: consumer spaced repetition is a UX product. The algorithm is prior art.
Written by Brainback Research. Published May 5, 2026. Filed under Method.