Day 28, 192 cards, and the day the queue peaks

Sacar una cita, to make an appointment, enters a deck on day 1. A tutor said it once in passing while rescheduling. Graded Good, it returns that same evening, then on day 3, then on day 14. On day 14 it comes back wrong, as hacer una cita, close and wrong. Graded Again, it reappears later that evening, then on day 16, then day 21, then day 33.

Seven touches in a month, well under a minute of total attention. Those days are not estimates. They come out of ts-fsrs 5.4.1, the scheduler library the app runs, at the app's shipped setting of 90% desired retention. Install the library, grade one card Good until a single failure on day 14, and you will get the same seven dates.

Week one: the deck fills faster than you expect

You record a lesson, or import audio you already have. A 50-minute tutor session transcribes and gets mined for both single words and multi-word phrases, so acabar and acabar de arrive as separate candidates.

An hour of real conversation usually surfaces a few dozen terms that were new or shaky for you. The exact count depends on your level and your tutor, and the app does not cap it.

Keeping all of them is a mistake. You tap through the list and select. Keep 10 to 15 from a first lesson. Selection takes about three minutes. Keeping fewer is the correct instinct: a card you did not want is a review you will resent on day 18.

Each kept term becomes a card with an example sentence drawn from the way you actually met the word, an image, and two recordings of that sentence. One recording replaces the answer word with a tone, so you can be tested by ear. The other is the complete read.

Then a detail that changes the arithmetic: each word is scheduled as two separate cards. A recognition card shows the Spanish and asks for the meaning. A recall card shows the meaning and asks you to produce the Spanish. They are scheduled independently, because recognizing acabar and producing it are genuinely different states of knowing.

So 12 kept words from one lesson is 24 scheduled cards. Two lessons in week one is 48. Your first sessions run longer than you expect, because new cards come back the same day and again the next day.

Week one has no payoff. Nothing has had time to be forgotten and recovered.

Week two: the intervals start doing the work

The scheduler is FSRS-6, which runs on your device. It tracks three things per card. Retrievability is the probability you can recall the card right now. Stability is the interval at which that probability has fallen to 90%. Difficulty is how hard the card is for you specifically, on a scale of 1 to 10.

Those three numbers are computed from 21 parameters fitted against review histories, and the target is set to 90% retention. Anki's manual, describing the same default, says it "should normally translate to remembering around 90% cards when they are reviewed, and only failing around 10%." That failure rate is the setting working as designed. A day where you pass everything means your intervals are shorter than they need to be.

What you see in week two: cards you passed on day 1 come back on day 3, then day 14. A word you graded Again comes back the same evening instead. The queue stops growing in step with your capture rate, because well-known cards leave the near-term queue.

Grade a card Again whenever the answer arrived after you flipped it. Recognizing the answer once you see it is a different event from producing it, and the scheduler has no way to know which one happened unless you tell it honestly.

Week three: the first real test of the method

Around day 14 to day 20, the cards from week one start returning at intervals of two weeks and up. This is the first moment the deck tells you something you had no other way of finding out.

You will pass most of them. You will also fail some words you were sure you knew, and those failures are the interesting ones. They are usually words you met once, in a sentence where context did most of the lifting.

There is a documented reason to expect this. Karpicke, Butler and Roediger surveyed 177 college students in Memory in 2009 about how they study. Most of them reread their notes and textbooks. Few of them tested themselves. Asked directly whether they would reread a chapter or try to recall it, the majority chose rereading. The authors read this as students experiencing a sense of competence while studying that the material does not support.

Your sense of which words you know is close to worthless. The deck gives you the real answer, card by card, without you having to arrange a test.

The uneven spacing in that sacar una cita sequence is the point. Cepeda and colleagues (2006) reviewed 317 experiments on distributed practice and found that the gap producing the best retention grows as the delay before the test grows. A fixed weekly review cannot do that. A scheduler fitted to your own failures can.

Week four: the shape of the queue

Here is a full 28-day run. I do not have a logged deck from a real user to publish, and Verbamor publishes no per-user averages, so I simulated one against ts-fsrs 5.4.1 at 90% retention, the same library and setting the app ships. Two lessons a week, 12 kept words each. Every card is graded Good, except when a coin weighted by the scheduler's own forecast for that exact review, its calculated probability of recall, the same number that set the interval in the first place, comes up against it. That is what keeps the failures honest: they come from the model's own retrievability, not from a number I picked.

The run produces 192 scheduled cards from 96 words by day 28. The daily count does not climb steadily, and it does not level off early either. The busiest day is day 24, at 59 cards, well into the second half of the month, once several lessons' worth of cards have matured into overlapping review windows. Day 17 comes close, at 57. The month's median, counting only the days something is actually due, is 32 cards.

Reviews land on 23 of the 28 days. The other 5, mostly in the first two weeks, are empty, because nothing has come due yet. That is the most useful thing in the run: the load is lumpy and it grows, not a flat ten minutes every evening from day one.

By day 28, 20 of the 24 cards captured on day 1 are scheduled 30 or more days out. They have left the month entirely. This is what compounding looks like in a deck, and it is the reason the daily count stops tracking your capture rate.

If a 32-card median sounds like more than you signed up for, capture less. At 8 kept words a lesson the same run gives 128 cards, a median study day of 19, and reviews on 24 days out of 28. The lever is the number you keep per lesson, and it is the only lever that matters.

Where this does not work

Skip this method if you are not generating your own input. The whole mechanism depends on having real lessons, conversations or audio to mine. Somebody with no tutor, no class and no listening habit gets a thin deck and would be better served by a course with content built in.

Skip it if you want fluency without daily contact. A schedule that assumes you show up is worse than useless if you show up twice a week, because cards pile past their due date and the intervals stop meaning anything.

Skip it if you are studying a language outside the four Verbamor teaches, which are Spanish, French, Italian and Portuguese.

And a limit worth stating plainly: capture and review builds words you can produce on demand. The ability to hold a conversation at speed comes from conversation itself. The deck exists so the words you meet in one conversation survive until the next one.

Three checks you can run on day 28, each producing a number rather than a feeling. Open the Library and read your week-one words: count how many you can put into a full sentence out loud, not just translate. Look at your daily card count over the last seven days: if it is climbing steeply, keep fewer words per lesson. Pick three captured phrases, not single words, and use them in your next tutor session.

The 20 minutes that start it

Record your next lesson, or import a recording you already have. Select 12 words, not 40. Study the cards that evening, because Gais, Lucas and Born (2006) found that vocabulary learned within a few hours of sleep was recalled better 48 hours later, and the deficit in the sleep-deprived group persisted even after a recovery night.

Then open the app the next day, and the day after.

One measurement to aim at: on day 28, produce 8 out of 10 of the words you captured on day 1, cold, from English. It is checkable in four minutes, and it is the difference between a deck and a list.

On the numbers: the named studies are published research, checkable by title and year. The seven-touch card trace is exactly reproducible: install ts-fsrs 5.4.1, run those seven grades at 90% desired retention, and you land on those seven dates. The 28-day run uses the same library and setting, but its failures come from a coin weighted by each card's own retrievability, so a second run will not reproduce the same 59-card peak. Neither run is a measured average across Verbamor users.