FSRS is the current state of the art for flashcard scheduling, and if you are building a flashcard app you should probably use it. We are not, and the differences matter more than they first appear.
What a card assumes that a question breaks
- A card has one fact. A four-option question tests discrimination between four things. Getting it right means you separated them this time, against these particular distractors.
- A card cannot be guessed. A four-option question is correct one time in four by chance alone. Feeding a lucky guess into a scheduler as a success pushes the interval out on material you do not know.
- A card is self-graded. You say how well you recalled it, after seeing the answer — which is exactly when people are least accurate about themselves.
- A card is reviewed in isolation. Exam questions arrive mixed, and the mixing is part of the difficulty.
Treating a one-in-four guess as evidence of knowledge is the single biggest way a card scheduler misleads you on a question bank.
What we grade on instead
Two signals, both captured at the moment of answering rather than after: whether you were correct, and how confident you were before you submitted. That pairing separates the four cases a bare correct/incorrect cannot — knew it, guessed it and got lucky, was sure and wrong, knew you were shaky.
Each of those deserves a different interval. A confident correct answer can wait a long time. A confident error should come back soon, because you are carrying an active wrong belief. A lucky guess should come back as though you had got it wrong, because in every sense that matters you did.
Why confidence before, not after
Asking how well you recalled something after revealing the answer produces hindsight, not a measurement. Everything looks more familiar once you have seen it. Asking before you submit costs a second, is mildly uncomfortable, and is the only version that carries information.
The self-grading problem in detail
Flashcard schedulers ask you, after seeing the answer, how well you recalled it. That question is asked at the worst possible moment: the answer is in front of you, everything looks familiar, and human judgement about one's own memory is poor under exactly those conditions. People systematically over-rate recall they have just been shown.
Asking before the reveal costs the same second and is a measurement rather than a reconstruction. It is a small design difference with a large effect on what the resulting data means.
What we do not claim
Ours is not a better algorithm than FSRS. On flashcards, with self-graded recall and no guessing, FSRS would beat it comfortably, and we would use it. The argument is narrower: for four-option exam questions, the inputs FSRS is tuned for are not the inputs available, and feeding it a one-in-four guess as a success does damage no amount of algorithmic sophistication repairs.
The best-tuned scheduler in the world cannot recover from being told a lucky guess was knowledge.
Interpretability is a feature
A scheduler that can say "this is here because you were sure and wrong three weeks ago" is one you can argue with, and being able to argue with it is what stops you abandoning it the first time it feels wrong. An opaque score has no answer to "why this question?" beyond asking for trust, and trust is what people withdraw when a tool tells them to study something they are confident about.
This will change
The honest position is that this is a first version tuned by judgement rather than by evidence, because the evidence does not exist yet at any useful scale. As attempt data accumulates it should be possible to check whether the intervals are right and adjust them. Anyone claiming a finely optimised schedule for exam questions today is claiming more than the data supports.
The honest limitation
A simpler, more interpretable scheduler will be beaten on raw efficiency by a well-tuned FSRS on material that suits cards. We think that trade is right here, because a scheduler you can explain is one a candidate can trust and correct. Due Today is live, and it tells you why something is due rather than asking you to take it on faith.