Roadmap
What comes next follows the same rule as the rest of the project: measure before deciding. Nothing goes in because it looks useful; it goes in when the data shows it is missing.
Next
- Tuning with real data. After two weeks of use, compare closed-book tests with practice with help before touching any parameter (the overconfidence threshold, the weak-foundation factor, the 10% floor).
- Reminders for the weekly closed-book test and the day before an exam (PWA notifications).
Phase B — question bank
The current version works at session level: you know what you did, how confident you were and what help you used. Phase B takes the rules down to item level (each question or exercise):
| Rule | With items |
|---|---|
| R1 attempt before revealing | A single point on the server decides whether an item's solution can be shown; only after an attempt has been logged in that session. Tested with mutations |
| R2 phased AI | Before the attempt, only hints (rephrasing, pointing to the concept, asking questions), with a leak check against the stored solution; afterwards, marking and explaining; in tests, nothing. hint_level_used per attempt |
| R3 spacing | FSRS per item, with reviews capped at each subject's exam date |
| R4 interleaving | The session builder never puts more than 2 items in a row from the same topic; "which method applies?" items |
| R6 calibration | Per topic, not just per subject |
The condition for moving ahead is content: an item bank needs correct questions for each subject. AI-generated items go in as pending and are only served once validated — zero items is better than wrong items.
Considered and set aside
- AI-generated summaries — low utility (Dunlosky et al., 2013).
- An AI tutor that gives answers — the crutch effect (Bastani et al.).
- Points, badges and leaderboards for hours — they would reward the wrong signal.
Got an idea? Open an issue.