Which memory model predicts my Danish best?
A spaced-repetition app schedules each card for the moment I am about to forget it. The champion model decides; challengers predict every review in shadow mode. This page shows, week by week, how well each one predicted what actually happened, and whether a challenger has earned the right to take over.
The report data could not be loaded.
The gate: is a challenger ready?
Week by week
Only reviews at least a day after the card's previous one count: minutes after learning every model predicts nearly 100%, which says nothing about memory. Weeks with fewer than such reviews show only their count.
Prediction error (log loss)
Per model, on the same reviews. Lower is better.
Predicted vs actual recall
What the champion expected, and what happened. A widening gap is drift.
Calibration, all reviews so far
When a model says 80%, are 80% of those cards remembered? On the diagonal is perfect.
Evidence per week
Reviews a day or more after the previous one: what the gate can learn from.
Models and promotions
Every model version is registered with its exact parameters, so a library upgrade cannot silently change a live model. A promotion rebuilds every card's schedule from the append-only review log.
How to read this
Log loss
The main score. It punishes confident mistakes: saying 95% for a card I then forgot costs far more than saying 60%.
Calibration
The scheduler acts on the probability itself (a card is due when recall would drop below 90%), so a model must be right about the number, not just the ranking.
Bootstrap over cards
Reviews of the same card are not independent. The gate resamples whole cards and asks how often the challenger still wins; one lucky week is not enough.
Why "no promotion" is a result
With a few dozen usable reviews a day, the evidence takes weeks. Until then the champion keeps scheduling; the gate refusing is the system working.