Numbers you can check

Every flashcard you review is scheduled by FSRS-5, the modern successor to the SuperMemo family of spaced-repetition algorithms. The model maintains a stability estimate for each card and from it a retrievability: the probability you would recall the card if asked right now. That probability decays between reviews along the FSRS forgetting curve and jumps when you review.

Everything called mastery in Chameleon is this retrievability, averaged over the cards behind a concept and bucketed into green, yellow and red at fixed thresholds. It is a model of memory with published behaviour, so when a number surprises you, there is a definition to check it against.

The readiness figure is honest about what it is: your memory state, weighted by what matters. Each knowledge atom in a plan carries a criticality (core, extended, peripheral) and an exam relevance assigned during generation. Readiness weights each atom's recall probability by both, so ten peripheral facts you know cold cannot mask a core mechanism you have let slip.

Both numbers appear together, with the delta, so you can see how the weighting changed the story. What readiness is not: it does not yet fold your quiz scores or written answers into the figure itself. Those are measured and shown alongside it, and we would rather show you two honest numbers than one blended one we cannot defend.

Because every card has a decay curve, the model can run time forward. The forgetting forecast projects each currently-strong card to the days ahead and shows, day by day, how much of your material falls below the recall threshold if you stop reviewing, and which concepts fall first. With an exam date set, the projection runs to exam day, which turns “I should review at some point” into “these three concepts will not survive until Thursday.”

Every review is also a test of the model itself. Before you answer, the scheduler has a prediction of how likely you were to recall the card; after you answer, the outcome is recorded. The calibration panel bins those predictions and compares them with what actually happened, and summarises the gap with the same error metric the FSRS benchmark literature uses.

This is measurement, stated plainly: the panel shows you where the model over- or under-estimates your memory, and tracks it week by week. It does not silently adjust your scores. If our forecasts are off for you, the place that says so is one screen down from the forecast.

Multiple choice tests recognition; a blank page tests production. Chameleon separates your accuracy by question type, because a comfortable multiple-choice score routinely hides a much weaker open-ended one, and most exams are the second kind.

Written answers are marked by the AI persona on a ten-point scale with two or three sentences of specific feedback, and those grades feed the same panel. So the split you see is between real graded production and keyed recognition, on the same material.

The insights page carries the rest of what the record supports: a coverage audit that lists the core material you have never tested at all, accuracy by hour of day and by session length, an interleaving score across your plans, and per-source mastery when a plan is built from several files. Each panel appears only when there is enough data behind it to mean something.

All of it flows into a print-ready report, laid out for paper, that you can save as a PDF from your browser and hand to a tutor, a study group, or your future self the night before.

Analytics that only report are a scoreboard. The same mastery engine that draws these panels feeds the reader: concepts your record marks as weak are flagged in the chapter spine with a revisit note in the margin, and their prediction prompts open ready to test you again. The measurement and the teaching run on one ledger.