How studying works

Predicted test score

Your predicted test score is the estimated score if a test covered everything in the set today. Every concept counts — not only the ones you have seen.

Seen concepts use their decayed recall right now. Unseen concepts use a prior learned from your first-attempt accuracy on this set, shrunk toward a conservative default until you have about twenty first attempts. That prior is personal to you and this set; it is not a claim that you know nothing of the material you have not touched.

The model prices real exam noise: a 20% chance of guessing right when you do not know an item, and a 10% chance of a careless slip when you do. A fresh, untouched set starts near 34% — the guessing baseline — and the score caps near 90% (perfect recall, still some slips). Skipping material does not remove it from the score — unseen concepts stay priced at your prior until you cover them.

Grading

Multiple-choice and true/false answers are graded on our servers against the stored answer key for that question. Free-text answers are graded by the model against the concept. Estimated recall and your predicted test score update from that server (or model) verdict — not from a local guess on the device.

Estimated recall and decay

For every concept, we track the probability that you know it right now. We call this $P_{\text{known}, t}$. After you leave a concept alone, that probability decays because memory fades without practice:

• $P_{\text{known}}$: how well you knew the concept at the last review.
• $f = 8 \times 10^{-7}$: how quickly estimated recall decays over time.
• $\Delta t$: seconds since the last review.
• $S_{\text{memory}}$: memory strength — higher means slower decay.

Concepts you have never reviewed start at $P_{\text{known}, t} = 0.2$.

Memory strength grows when you succeed on a well-timed review: the harder the retrieval (lower estimated recall at the moment you answer), the bigger the gain — discounted when a lucky guess was likely. It drops when you miss.

When a concept is due

A concept is due when its estimated recall falls below 75%. Due concepts are the ones worth reviewing first; new (unseen) concepts fill the rest of a daily plan.

If you set an exam date on a study set, the plan paces how many new concepts to introduce per day so the set can be covered before the exam, with a small buffer. Without an exam date, a modest default number of new concepts is suggested each day.

What gets asked next

Free-response study picks one concept at a time. Score is $(1 - P_{\text{known}, t}) \times \text{importance}$: lower estimated recall and higher importance rank first. Concepts from the same section as your most recent answer get a 1.2× boost when they are new or still weak, so related material tends to stick together. Recent answered concepts are excluded — up to the last five tests — so the next question is not a repeat of what you just saw.

When any section's estimated recall is below 60%, free-form study locks onto the weakest section until that section recovers above the threshold.

Quick-quiz batches (multiple choice and true/false) use the same ranking idea: concepts ordered by $(1 - P_{\text{known}, t}) \times \text{importance}$, then a short batch is generated for those top concepts.

Why this shape

The system combines spaced repetition (review before estimated recall slips too far) with importance weighting and light section cohesion. The headline predicted test score uses that same recall math on seen material, plus a personal prior on unseen material and exam-style guessing and slips — so the number you study toward is the one a full-set test would estimate today.