The science behind Peabody

What this page is

Peabody (peabody.kids) is a free AI-driven adaptive learning platform for kids, and every rule that adapts is written down here with the research it draws on and the limits of that research. AI writes the stories. Plain, checkable code decides the level, the next question and when something comes back. The rules fit in a few sentences each, and this page is the long version, kept honest: Peabody has not been tested in a trial, and nothing below is a promise about outcomes. It is free through the end of 2026.

Try it first: start free or read a story from the library.

Two jobs, two kinds of tool

Teaching a child through software is two very different jobs. One is making something worth doing: a story a child wants to read, a game they would pick anyway. Language models are good at the first and Peabody uses one (Claude, from Anthropic) for stories, for safety screening and for reviewing a sample of stories for quality. The other job is deciding what fits and remembering what happened: which level, which question, when to repeat. That needs the same answer every time and a rule a parent or teacher can read. That is ordinary code. A child's level is never a language model's opinion, and every AI call is budgeted and logged per child.

The rules, each with its evidence

1. Known things come back after 1, 3, 7 and 21 days

Spreading practice over several days produces better long-term memory than the same amount of practice in one sitting. This is one of the most replicated findings in memory research (Cepeda and colleagues' 2006 review of 254 studies; their 2008 study of the best gap between sessions). Smaller studies of children learning word meanings in grades 3 and 5 found the same pattern (Sobel, Cepeda and Kapler 2011; Goossens and colleagues 2012 and 2016).

In Peabody a fact or word a child answered quickly and right comes back for review after 1 day, then 3, then 7, then 21, the gap growing with each separate day it is right. A miss brings it back sooner. For the practice words in a story, a word counts as known only after three right cold reads on three different days, with keeper checks at a week, a month and a quarter.

Limits. The benefit is smaller for younger children and short delays, and spacing has not been shown to raise overall reading ability. The number three for the day rule is a design choice supported by small correlational studies of spoken word learning (Gordon and Lowry 2024), not a tested threshold for printed words.

2. A cold try first, then the answer

Trying to recall something before seeing it usually strengthens memory more than looking at it again. The effect is strong in adults (Rowland's 2014 meta-analysis) and shows up in children from preschool on when the recall is cued, feedback is immediate and the child can mostly succeed (Fazio and Marsh 2019). Every story starts with a short warm-up where the child says whether they can read each practice word, with no hint. Every game question is a cold try with the answer shown at once. In the child's view it is a game, never a test.

Limits. One multi-grade classroom vocabulary study found no benefit over restudying, and no study has tested retrieval practice for learning to read printed words in a home app.

3. Most answers are wins, by design

A round is about 30 questions, and about 85% of them are things the child already knows, with at most two new items taught first and rehearsed a few questions later. A miss comes back three questions later so the round ends on a success. Two misses in the last five and no more new items this round; three and the game eases. The idea that learning goes best when a learner succeeds most of the time but not always has support in a 2019 analysis of learning systems that found an optimal error rate near 15% (Wilson and colleagues, Nature Communications), and in the long-standing practice of keeping practice "winnable" (Rawson and Dunlosky's successive relearning work).

Limits. The 85% figure comes from models of learning machines and adult tasks. For a five-year-old it is a sensible default, not a measured optimum.

4. Every subject is a ladder, and a grade only sets the first rung

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 rung is passed when enough of it has been seen, 80% of what was seen is solid, and only a few items are still being learnt; a child who has seen a dozen questions on a rung and got 88% of every answer right tests out of it. Passed rungs apply in every game at once. This is mastery learning: move on when the skill is solid, not when the calendar says so. Kulik, Kulik and Bangert-Drowns' 1990 meta-analysis of mastery learning programs found moderate positive effects, and the orders themselves follow published learning trajectories for early math (Clements and Sarama) and a hand-written phonics scope and sequence.

Limits. Bloom's famous "two sigma" result for tutoring plus mastery learning has not held up at that size in later trials. The rung sizes and the 80% bar are Peabody's own calls, and are published so they can be argued with.

5. A question right once is never asked again

Once a child has answered a question right, first try, Peabody retires it and moves the skill on to a fresh question, not a fresh lesson: when "fish" is done, another picture that starts with f comes in as known. A fact retires one way of asking at a time (3 + 4, 4 + 3 and 3 + ? = 7 are separate questions). A question is asked at most twice: a miss comes back once, later in the round, and is then done whichever way it goes. This follows the finding that extra practice on an item already learnt in the same session pays little (Rohrer and Taylor 2006 on overlearning), and it keeps a child who is past adding from seeing adding everywhere. Before this rule, a simulated kindergartner who got everything right saw 374 of 436 questions they had already answered right twice. The rule began as right twice; since 16 October 2026 one right is enough, because the second right in the same sitting added little.

Limits. Retiring a question is not the same as remembering it a month later. That is what the review schedule in rule 1 is for. With three boxes a guess is right one time in three, so one right can retire a question that was not learnt, and a word missed twice is gone rather than learnt; in both cases the skill still comes back through fresh questions, and passing a rung still needs most answers right.

6. Speed is a signal, but never a clock on something new

Answering a math fact in under about three seconds is a common marker of automaticity in the fact-fluency literature, and Peabody records time to answer on every question: 3 seconds for adding and taking away, 2.5 for times and dividing, 2.5 to hear a word, 3 to decode one. A quick first-try right answer is what moves an item up the levels. New items never have a timer, and a wrong answer costs tempo in a game, never a life.

Limits. Timing on a tablet includes finding the box with a finger, so thresholds are generous, and a parent sees "fast in games" kept separate from "known in reading".

7. The learning lives inside the game

Meta-analyses of game-based learning find that games beat ordinary instruction on average and that the design matters more than the medium (Clark, Tanner-Smith and Killingsworth 2016; Wouters and colleagues 2013). One design finding in particular shapes Peabody: when the learning mechanic is the core of the game rather than a quiz bolted between levels, children learn more and choose to play longer (Habgood and Ainsworth 2011, who tested it with a math game). So in Peabody the question sits on a thing in the world, the answer is one of the game's own moves, and runners come to a full stop at a station rather than slowing while a child reads. Every game also has a Just play mode with no questions, because a game that only works as a quiz is not a game.

Limits. Game studies are mostly short, often in classrooms, and transfer from a game to a classroom skill is uneven. Peabody measures how long kids choose to play each game and retires the ones they drop.

8. A story's level is measured in the text, not guessed about the child

Each story and library book is measured with sentence length, Flesch–Kincaid, the share of words off the Dale–Chall familiar list, words per screen, and for early readers the share of sight words and how decodable the words are at the child's phonics step. A child picks where to start by tapping the sentence they can read comfortably, and the level moves when their reading says the books ran too easy or too hard. The writer is told to aim under the level's middle, since a model writes harder than it is asked, and a story over its ceiling is repaired or rewritten.

Limits. Any readability formula wanders by about a grade on a short text, so every story gets a range. Peabody doesn't measure Lexile, doesn't listen to a child, and so cannot measure fluency or decoding. It is not an assessment.

9. The reward for reading is the next story

Self-determination theory (Ryan and Deci) holds that motivation lasts when people feel they chose the activity, can succeed at it, and do it with people who matter to them. A child picks the story's ideas and mood and which game to play; most answers are wins; and stories are made to be read to someone. Deci, Koestner and Ryan's 1999 meta-analysis found that rewards given for doing an activity can undercut interest in it, so Peabody's coins come from the game's own world and learning modes simply pay more.

Limits. A large field trial (Fryer 2011) found incentives for inputs like reading books did help, so "rewards hurt" is not a law. Peabody keeps rewards small and tied to doing, not to being better than someone.

10. What Peabody refuses, and why

No leaderboards, no percentiles, no comparison to other children, no countdown timers, no loss of what was earned, no random reward boxes outside a game's own eggs, no guilt-tripping mascot, no open-ended chat with an AI. Comparison feedback shifts a child's attention from the task to themselves (Butler 1988), and a 2022 study of apps used by children found manipulative design in most of them (Radesky and colleagues, JAMA Network Open). The streak is a gentle glow that comes back the next day.

What Peabody has not shown

• Peabody has not been tested in a trial. No claim on this site says it improves reading or math.
• The spacing and retrieval studies above measured word lists and facts, not reading level.
• "Adaptive" here means specific rules that change specific things because of what a child did. It does not mean the software understands a child.
• If reading or math worries you, the right first step is a conversation with your child's teacher, and a professional assessment if they suggest one.

How to check any of this

Every rule is a number in the code and a line in a test. The day gaps (1, 3, 7, 21), the two-rights rule, the 80% pass bar, the 85% known share, the speed thresholds and the ladders are checked by a test suite that runs before any change to a game can land, and the founder's label is needed to change the test itself. Parents with a free account see the same data in plain words: which words and facts are known, which are being learnt, which were slow, and one thing to try tonight.

Read more: how Peabody tracks words, where is my child stuck, and what reading level means.

Questions parents ask

Is Peabody evidence-based?

Its rules draw on published research on spacing, retrieval practice, mastery learning and game design, and each rule above names the studies and their limits. Peabody itself has not been tested in a trial, so it does not claim to raise scores.

What does the AI actually decide?

Nothing about your child's level or progress. The AI writes stories and screens them for safety. Code decides the level, the next question, the review schedule and what you see.

How is this different from a quiz app?

Most answers are things your child already knows, new things are taught before they are asked, a question answered right never comes back, and in the games the question is part of the world and the answer is a move. Every game also plays whole with no questions at all.

Can I see the rules for my own child?

Yes. A free grown-up account shows which words and facts are known, learning, missed or slow, which books ran too easy or too hard, and one thing to practise tonight. Nothing is a score, and nothing compares your child to anyone else.

Does it diagnose dyslexia or a learning difficulty?

No. Peabody never listens to a child and does not give a Lexile or any norm-referenced score. Talk to your child's teacher.

References

• Cepeda, Pashler, Vul, Wixted and Rohrer (2006). Distributed practice in verbal recall tasks. Psychological Bulletin. escholarship.org
• Cepeda, Vul, Rohrer, Wixted and Pashler (2008). Spacing effects in learning: a temporal ridgeline. Psychological Science. doi.org/10.1111/j.1467-9280.2008.02209.x
• Sobel, Cepeda and Kapler (2011). Spacing effects in real-world classroom vocabulary learning. Applied Cognitive Psychology. yorku.ca
• Goossens et al. (2016). Distributed practice and retrieval practice in primary school vocabulary learning. Applied Cognitive Psychology. doi.org/10.1002/acp.3245
• Gordon and Lowry (2024). Number of sessions children retrieve words relates to retention. Journal of Child Language. pubmed
• Rowland (2014). The effect of testing versus restudy on retention: a meta-analytic review. Psychological Bulletin. pubmed
• Fazio and Marsh (2019). Retrieval-based learning in children. Current Directions in Psychological Science. sagepub
• Wilson, Shenhav, Straccia and Cohen (2019). The eighty five percent rule for optimal learning. Nature Communications. nature.com
• Rawson and Dunlosky (2022). Successive relearning. Current Directions in Psychological Science. sagepub
• Kulik, Kulik and Bangert-Drowns (1990). Effectiveness of mastery learning programs: a meta-analysis. Review of Educational Research. sagepub
• Clements and Sarama (2020). Learning and Teaching Early Math: The Learning Trajectories Approach. Routledge. learningtrajectories.org
• Rohrer and Taylor (2006). The effects of overlearning and distributed practise on the retention of mathematics knowledge. Applied Cognitive Psychology. doi.org/10.1002/acp.1266
• Clark, Tanner-Smith and Killingsworth (2016). Digital games, design, and learning: a systematic review and meta-analysis. Review of Educational Research. pmc
• Wouters, van Nimwegen, van Oostendorp and van der Spek (2013). A meta-analysis of the cognitive and motivational effects of serious games. Journal of Educational Psychology. doi.org/10.1037/a0031311
• Habgood and Ainsworth (2011). Motivating children to learn effectively: exploring the value of intrinsic integration in educational games. Journal of the Learning Sciences. doi.org/10.1080/10508406.2010.508029
• Ryan and Deci (2000). Self-determination theory and the facilitation of intrinsic motivation. American Psychologist. selfdeterminationtheory.org
• Deci, Koestner and Ryan (1999). A meta-analytic review of experiments examining the effects of extrinsic rewards on intrinsic motivation. Psychological Bulletin. doi.org/10.1037/0033-2909.125.6.627
• Fryer (2011). Financial incentives and student achievement: evidence from randomized trials. Quarterly Journal of Economics. nber.org
• Butler (1988). Enhancing and undermining intrinsic motivation: the effects of task-involving and ego-involving evaluation. British Journal of Educational Psychology. eric.ed.gov
• Radesky et al. (2022). Prevalence and characteristics of manipulative design in mobile applications used by children. JAMA Network Open. jamanetwork.com