How Peabody picks the next question: the practice brain, rule by rule

Every question a child sees in a Peabody game comes from one small engine, called the practice brain. It is ordinary code, not AI, and this guide walks through it rule by rule with the real numbers: six item levels, a review schedule of 1, 3, 7 and 21 days, decks that are about 85% things the child already knows, a mastery ladder per subject, a test-out rule, "a question right once is never asked again", and stretch items for a child who is cruising. Peabody (peabody.kids) is a free AI-driven adaptive learning platform for kids, free through the end of 2026, and the point of writing this down is that a parent or teacher should be able to read the whole rulebook in ten minutes.

If you want the research behind each rule, read the science behind Peabody. This guide is about the mechanism.

The shape of it

A strand is a skill. Math has seven: counting, making ten, adding, taking away, number sense (the number line to 100, skip counting by 2, 5 and 10, comparing, place value), times tables and dividing. Reading has six: first sounds, rhymes, phonics, sight words, spelling and word meanings.

An item is one thing to know inside a strand: the fact 3 + 4, the sight word said, the phonics word ship, the first sound of fish. Each child has a small record per item. A question is one way of asking about an item. A fact has several (3 + 4, 4 + 3, 3 + ? = 7); a word item has one.

A deck is one round of about 30 questions, built fresh each time a game asks for one. A round is a deck played through in a game, with the answers sent back as they happen.

Everything below is how the brain fills a deck and what it does with the answers.

Rule 1: six item levels, moved by every answer

Each item sits at a level from 0 to 5 for that child:

The levels only ever use what the child did. A wrong answer sends an item to level 1 and resets its count of quick days. A hint or a retry caps it at 2. A plain right answer is 3. A quick right answer is 4, and a second quick right answer on a different day makes it 5.

"Quick" is a fixed threshold per strand, measured from the moment the prompt appears to the moment the child commits an answer: 3 seconds for adding, taking away, counting and making ten; 2.5 seconds for times tables and dividing; 6 seconds for number-sense questions; 2.5 seconds to hear a word and pick it; 3 seconds to decode a printed word; 2 seconds plus 1 second per letter to spell; 5 seconds for a word meaning. The thresholds are deliberately generous because they include finding the box with a finger on a tablet.

One kind of question moves no level at all: a probe. About once every 40 answers the brain slips in a measurement-only question so that it has an unbiased reading of what a child knows, separate from the questions chosen to be winnable.

Rule 2: known things come back after 1, 3, 7 and 21 days

Once an item is at level 3 or above it is a known item, and it gets a due date. The gap grows with the number of separate days the child has answered it quickly and right since the last miss: 1 day after the first, then 3, then 7, then 21. A miss drops the item to level 1 and the ladder of gaps starts again.

This is the spacing schedule, and it is the same for every child. It is a Leitner-style ladder rather than a model that fits a forgetting curve per child, on purpose: the gaps can be read off a table, and a change to them is a change to one line of code and one line of a test.

The practice words inside a story use a separate, stricter ladder, described in how Peabody tracks words: a word counts as known only after three right cold reads on three different days, with keeper checks a week, a month and a quarter later.

Rule 3: a deck is about 85% wins

A deck is about 30 questions, filled in this order:

1. Due reviews first. Known items whose due date has arrived, the most overdue first.
2. Learning items. Items at level 1 or 2 that the child is still working on.
3. New items, at most two a deck and four a day per domain (math or reading). Each is taught first, with a Teach moment that shows the answer, and comes back a few questions later to be tried.
4. Known items, longest unseen first, to fill the rest, with at least ten distinct ones so a round never loops on two or three.

The target is that about 85% of a deck is things the child already knows. The research on an optimal error rate (around 15% in a 2019 analysis of learning systems) is one reason; the other is that a five-year-old who is losing most of the time stops playing.

Inside a round, a missed question comes back three questions later so that the round ends on a success. There is a frustration guard: two misses in the last five answers and the round adds no more new items; three misses and the game is told to ease, which in a runner means a longer stretch of plain running before the next station.

A right first-try answer with no hint pays 1. A right answer after a miss or a hint pays a quarter. A wrong answer pays nothing but costs nothing either. Wrong answers cost tempo in a game, never a life.

Rule 4: every subject is a ladder, and a grade only sets the first rung

The strands are ordered. Math climbs counting → making ten → adding → taking away → number sense → times tables → dividing. Reading climbs first sounds → rhymes → phonics → sight words → spelling → word meanings.

A grade, if a child or a grown-up picked one, only decides the first rung: kindergarten starts at counting and first sounds, first and second grade at adding and phonics, third grade and up at times tables and word meanings. From there the brain decides.

A rung is passed when three things are true:

• Enough of it has been seen. The bar is per strand: 8 counting items, 8 making-ten items, 24 sums, 24 subtractions, 10 number-sense items, 30 times facts, 24 division facts, 20 first sounds, 16 rhymes, 30 phonics words, 30 sight words, 20 spellings, 30 word meanings. For a short rung the bar is capped at 60% of the rung.
• 80% of what was seen is solid: at level 3 or higher, or done for good, or right at least twice with the last answer right and three rights for every miss.
• Only a few items are still shaky: wrong the last time, or wrong more often than not. Two shaky items, or one in ten, is the most a passed rung may have.

There is also a test-out rule, added after a real child who was past adding kept seeing sums everywhere. A rung with eight or more questions seen and 88% of every answer right is passed, whatever the day-by-day levels say.

Rounds rotate between the lowest two rungs not yet passed, so the next skill is always in view, and the next rung comes in once the child is halfway up the current one. About one round in four goes back over a passed rung for review. A child who misses a lot stays where they are. Passed rungs are saved per child and apply in every game and on every device, so passing adding in Rocket Number Blast means no more adding in Onslaught either.

Rule 5: a question right once is never asked again

Once a child has answered a question right first time, the question is done for good. The brain never builds it into a deck again while any other question exists; the only exception is a brand-new strand so small that nothing else is left, and even then the oldest done question is the last resort.

A question is asked at most twice, ever. A miss comes back once, later in the same round, and after that second ask it is done whichever way it goes. A deck holds each question once, so a thin rung makes a shorter deck rather than a repeat. The trade-off, stated plainly: a word missed twice is gone rather than learnt. Its skill still comes back through its siblings (below), and a question missed and re-asked never counts as solid toward passing a rung.

For word items the question and the item are the same thing, so the item retires. For a fact, each way of asking retires separately: 3 + 4 can be done while 3 + ? = 7 is still live. A fact is done for good once every way of asking it is right, or once it is locked in: right four times, quick, and right the last time.

The skill moves on, not the lesson. When fish is done, another picture that starts with f (fan, fox, frog) comes in as a known question with no Teach moment, and it does not count against the day's cap on new items. The same holds for a rhyme family and a phonics step. Games that make their own questions (Stone Age's science, Pet Hatch's quests, the World trails) share the same done list on the device, so the rule holds there too.

This rule exists because the opposite is what most practice software does. In a simulation before it shipped, a kindergartner who got everything right saw 374 of 436 questions they had already answered right twice, and never reached adding. After it, 6 of 526. The rule began as right twice; on 16 October 2026 it became right once, because a second right in the same sitting told the brain little and a child who was well past a skill was still asked everything twice.

Rule 6: stretch items for a child who is cruising

If everything a child has seen in a strand is solid and nothing is shaky, they are cruising, and the next six items in the strand's teaching order come in as stretch items: untaught, ahead of the known ones, past the day's cap on new items. A deck too thin to fill without asking a question again also climbs, taking up to sixteen items in order, and the deck keeps rising until the child misses. That works as placement for a child the brain has never met.

Two limits keep the start-easy rule: a child new to a subject climbs no further than a grade up's easy end, and a child who is struggling (more than two shaky items) gets no stretch at all.

Where the order inside a strand comes from

The brain teaches items in a written order, not a random one.

• Adding goes +0, +1, doubles, make ten, +10, near doubles, +9, then the rest by size. Taking away is the inverse of each sum in the same order.
• Phonics follows a sixteen-step scope and sequence, each step decodable with the letters taught so far: s a t p i n; m d g o c k; ck e u r; h b f l ff ll ss; word families with j v w; sh ch th; wh ng nk qu x y z; blends at the start; blends at the end; silent e; vowel teams ee ea ai ay oa ow; ar or er ir ur; igh and y as in fly; -le and two-part words; un- and re-; -ing -ed -er -est. Dolch words are taught separately as heart words, so a phonics step never leans on an irregular word.
• Times tables start with the facts most children find easiest (×2, ×5, ×10) and work out from there.

A brand-new child is treated as already knowing the easy end of each order (+0 and +1, counting to three, the first three words of the first phonics step), so the first deck has something to be right about.

What the brain never does

• It never shows a word as the answer to "which word is this". Peabody says the word; the boxes carry the choices. Printing cat above three boxes that say cat, cot and cut is letter matching, not reading.
• It never puts a clock on a new item. Speed is measured, never demanded.
• It never asks for typing. Every answer is three big boxes, or a drag-to-match or fill-the-gap with at least three targets.
• It never ranks children against each other, and never shows a child a score.
• It never diagnoses. "Shaky" and "slow" are descriptions of answers on a tablet, not of a child. Peabody doesn't measure Lexile and cannot hear a child read.

How you can check it

The numbers in this guide are constants in the code, and a test suite called the learning invariants runs before any change to a game can land: it checks that the never-twice constant is 1 (right once is done), the pass share is at least 80%, the review gaps strictly increase, every ladder is in grade order with every strand on a rung, and that simulated children (a perfect kindergartner, a kindergartner at 60%, a second grader at 85%) never get an empty deck or a retired question while another is left. Changing the test needs the founder's own label.

With a free grown-up account you see the same engine from the outside: which words and facts are known, which are being learnt, which were slow, and one thing to try tonight. Or skip the account and start free on any browser.

Questions parents ask

Is the practice brain AI?

No. It is a few hundred lines of ordinary code with the numbers in this guide. AI (a language model) writes Peabody's stories and screens them for safety; it never decides a level, a question or a review date.

Why not a smarter model that learns each child's forgetting curve?

Because the gain is small for a five-year-old answering thirty questions in a game, and the cost is that nobody can read the rule. Peabody's schedule can be read off a table and changed in one line. If a per-child model ever earns its place, it will be measured against this one first.

What happens on a bad night?

Two misses in the last five answers stop new items for the round; three tell the game to ease. A miss sends the item back to level 1 and it comes back three questions later so the round ends on a win. Nothing is lost that was earned.

My child is past adding but keeps seeing sums. Why?

Usually because the rung was not yet passed on the day-by-day levels. The test-out rule (eight questions seen, 88% right) exists for exactly this, and passed rungs now apply in every game. If it still happens, the grade picked on the math page may be holding the starting rung down; the ladder climbs past it as the child plays.

Does the engine work without an account?

Yes. A visitor's device keeps the same levels, done list and climb locally. A free grown-up account adds the plain-words view and keeps progress across devices.