Spaced repetition done honestly

Spaced repetition is a good idea with a bad interface. Every app that ships it hides what the box does, then asks you to help it decide when to show the card again by rating a thing you haven’t earned yet. You tap “Good,” the box swallows the tap, and a week later a card shows up with no explanation of why today.

We kept the idea and threw out the self-rating. Here is what we do instead, and why it holds up.

§

What classic SRS asks

SM-2 is the algorithm under Anki, SuperMemo, and most of the descendants. It schedules the next review of a card by multiplying the current interval by an ease factor, and that ease factor moves up or down based on a rating the user picks after each card.

die Erfahrung
experience (feminine, -ung)
AgainHardGoodEasy
You rate yourself

The user sees the prompt, reveals the answer, then picks Again, Hard, Good, or Easy. That pick is the only signal the algorithm gets. The box computes a new interval, puts the card in a queue, and shows it to you next week, or in a month, or tomorrow, or in ten minutes.

§

Self-rating is the problem

The thing you are being asked to rate (did I know that?) is the same thing you are trying to learn. Brains are famously bad at this. Recognition feels like recall. Reading the answer out loud feels like remembering it. The fluency illusion is a real, measured cognitive effect, and the four buttons at the foot of every SRS card sit right on top of it.

There is also the evening problem. At 11pm on a Tuesday, Hard and Good collapse into the same button. People press Good so the queue shrinks faster. The queue does shrink faster. Weeks later the cards they called Good come back as holes, and the algorithm had no way to know the user lied to it, because the user lied to it on purpose.

An honest learning system measures. It does not ask.

§

What we measure instead

Every answer in GenauApp is a real exercise, not a self-report. An article picker is three buttons under a noun; you tap one. That tap is the measurement. A fill in the blank has one right answer and a typed guess; the match is the measurement. Sentence Builder is a sequence; the order you built is the measurement.

dieErfahrung
experience
derdiedas
App measured, correct

The app never asks how you felt about that answer. It already has the answer. “How confident were you?” is a worse question than “did you get it right?” and the right question is the one with data under it.

§

Why one answer is a weak signal

A single correct answer does not mean you know a topic. Ten correct answers means something. Our progress bar multiplies your accuracy by the number of attempts over ten, capped at one. Early on your ring moves slowly even when you are getting them right, because two data points do not get promoted to a “learned” signal. Later, one bad day does not crash the number either, because the sample is big enough to absorb it.

Vocab·
Grammar↑
Writing·
Reading↓

This is spaced repetition’s honest version of “mature card.” A card is mature when the data says it is. Not when a deadline passed, and not when the user tapped Easy on it twice in a row.

§

Wrong answers cost you

Measured answers cut both ways. A right answer pushes the ring up. A wrong answer takes ring points off, directly. No “effort trumps accuracy” forgiveness, no silent decay on a timer: the ring is a measurement, and measurements include the wrong ones.

This is the piece SM-2 cannot do. When you tap Again on an Anki card, the box reschedules the card. It does not lower a score, because there is no score to lower. Your “did I know this?” is the input, and the output is just another interval. The number that tracks your German lives in your head.

Alongside the ring sits a small arrow: up, stable, or down. The ring is your level today; the arrow is your direction right now. It compares your last three answers in that skill to your historical average. Ten points better and it points up. Ten points worse and it points down. Everything in between is stable, and “stable” is not a shrug (we wrote up the full mechanic here).

SM-2 can say “your next review is in four days.” It cannot say “your recent answers are shakier than they used to be,” and it cannot cost you points when you miss. We can do both.

§

The whole loop

The loop, in order. You open the app and the daily plan picks one of your four chips for you, usually the one whose arrow is pointing the wrong way. You work through the exercises. The app measures each answer. Right answers push the ring up, wrong answers cost you points, and the arrow next to the ring tells you which direction your recent data is heading. No rate yourself. No ease factor. No opaque interval calculation.

Spaced repetition was always a good idea. It just needed the user to stop grading their own paper.

Dein Deutsch, dein Workout.

Join the waitlist. We will let you know when GenauApp launches.

Back to GenauApp →