Space practice around the retention horizon, not a magic interval
Use the time until the knowledge should be available, plus observed retrieval, to plan the next encounter without pretending one schedule fits everyone.
Spacing means returning to material across separated learning episodes instead of concentrating every encounter in one block. The research literature supports a spacing effect, while also showing that the useful interval interacts with the desired retention interval. That is a reason to design a measurable schedule, not a reason to copy one universal sequence of day numbers.
Name the horizon first
Decide what the knowledge should do and when. A casual long-term category foundation, preparation for an official practice test, and recovery of a frequently missed relationship have different horizons. Write the horizon beside the item or category plan. Without it, an algorithm can optimize short-term ease while the learner expects durable availability.
| Input | Question | Caution |
|---|---|---|
| Retention horizon | When should this remain retrievable? | A date is a planning input, not a guarantee |
| Last independent attempt | What happened before feedback? | Immediate post-feedback success is not independent |
| Confidence calibration | Did confidence match the outcome? | Confidence alone is not mastery |
| Error pattern | Is the miss repeated or changing? | A bad cue may imitate forgetting |
| Coverage | Which categories receive no attempts? | Easy-item repetition can hide blind spots |
Let retrieval change the interval
A correct, well-calibrated response may justify a longer interval. A confident error deserves an earlier discriminating encounter. A low-confidence correct response might need another cue before expansion. An ambiguous item should leave the scheduler entirely until reviewed. The rule should be explicit and versioned so changes can be evaluated against delayed outcomes.
The schedule is a policy under evaluation. It should expose its version, inputs, and limits instead of presenting the next-review date as a scientific fact about one learner.
Protect category coverage
Item-level success can pull practice toward material already present in the deck. Add a separate coverage view for broad category families and relationships. Reserve part of a session for underrepresented areas, while keeping the selection rule visible. Coverage does not mean equal time for every label; it means the system can show where its opportunities and evidence are thin.
Evaluate a schedule with delayed outcomes
Compare schedule versions on the same eligible definition, horizon, cue quality, and content cohort. Report delayed first-attempt accuracy, calibration, lapse distribution, coverage, and burden. Avoid sampled-negative shortcuts when the real serving corpus is small enough for exact evaluation. A schedule that produces more reviews may improve recall while making the practice plan unsustainable; both outcomes matter.