How adaptive learning knows where your child is stuck

An adaptive learning app can tell where a child is stuck from three kinds of evidence: whether answers were right, how fast they came, and the pattern across the last ten or so tries at the same skill. Peabody (peabody.kids) records all three for every item, in ordinary code with published thresholds, and turns them into one plain sentence for a grown-up and one thing to try tonight, never a score, a label or a diagnosis. It is free through the end of 2026.

This post walks through the science of each signal, what it can and cannot support, the thresholds Peabody actually uses, and why the app will not try to read your child's feelings.

Right or wrong is half the evidence

Most practice software keeps one number per skill: the share of answers that were right. It is a thin record. A child who gets 3 + 4 right after staring at it for eight seconds and a child who taps 7 before the question has finished appearing both score the same, and they are in different places.

The field of student modeling has spent two decades adding channels to that one number. The two that work from a browser alone, with no camera and no microphone, are time and pattern.

Speed: strong evidence, easy to misread

Timing is the cheapest second channel. For math facts, school-facing material converges on "right within about 2 seconds" as the sign that a fact is retrieved rather than worked out, with 2 to 3 seconds in intervention studies and 3 seconds as the generous ceiling (Burns and colleagues 2010, as cited in a 2024 ERIC study; Rocket Math summary, 2016). Lab measures of direct retrieval are faster, 400 to 900 milliseconds from stimulus to response (ERIC, 2019). For sight words there is no validated cutoff at all; the nearest numbers come from a US patent reporting mean latencies around 572 ms for second graders and 937 ms for sixth graders, which is not peer-reviewed evidence.

The caution comes from Wang and Heffernan (Educational Data Mining, 2012): a short response time can mean the child knows it or is guessing; a long one can mean careful thought or a wandering mind (PDF). Later models keep speed and ability separate for that reason, adding a quick-or-slow indicator next to the knowledge estimate rather than folding it in (reviewed in arXiv 1801.07618, 2018). The practical reading is simple: speed counts as evidence of retrieval only after the child has already shown they can get the item right slowly. A fast right answer on first sight is more likely a lucky tap.

Peabody's thresholds, measured from the moment the prompt appears to the moment the child commits an answer:

They are generous on purpose: a five-year-old also has to find the box with a finger. And speed never counts on its own. Each item sits at a level from 0 to 5 per child: 3 is right first try but not quick; 4 is quick and right first try with three boxes on screen; 5 is quick, first try, three boxes, on two different days. A fact is locked in only after four quick rights. New items never have a timer at all; speed is measured, never demanded.

"Stuck" has a definition: the wheel-spinning rule

The most useful single signal for "where is my child stuck" comes from Beck and Gong (Artificial Intelligence in Education, 2013) and Gong and Beck (Learning at Scale, 2015). They defined mastery of a skill as three correct answers in a row, and "timely" mastery as reaching it within ten opportunities. A child who keeps practicing without getting there is wheel-spinning, and in Cognitive Tutor data roughly a quarter of student-skill pairs did (Gong and Beck, 2015). Early detectors in ASSISTments can see it coming from the first seven tries, using hint requests, prior correctness, practice count and time on the skill (ERIC, 2019).

The rule's strength is that it needs no model and no labeled data; a product with a hundred users can run it. Its weakness is that definitions disagree across datasets (ERIC, 2020), and a 2026 case study on MATHia argues most published definitions are out of step with how real systems actually stop practice (arXiv, 2026).

Peabody does not label wheel-spinning by that name yet. What it has is the same shape at the level of one item and one round. An item is shaky if it was wrong the last time or wrong more often than not; a rung cannot be passed with more than two shaky items or one in ten. Inside a round, a miss comes back three questions later; two misses in the last five answers stop new items for the round; three tell the game to ease. A child who misses a lot stays on the rung they are on. The three-in-a-row-within-ten rule is the natural next step, and it is the one this post would recommend to any product.

Rushing and gaming

The other pattern worth detecting is a child who is not trying: tapping fast and wrong, or clicking through hints to the answer. Baker, Corbett, Roll and Koedinger (User Modeling and User-Adapted Interaction, 2008) showed that detectors of this "gaming the system" transfer across student groups and, less well, across lessons (Springer). Paquette and colleagues (2015) found that hand-written rules ("three or more wrong in under a few seconds each") transfer between systems better than trained models do (PDF), and some gaming detectors still work on data collected sixteen years later (Levin and colleagues, 2022). The right response, in that literature, is a different activity, not a scolding.

Peabody takes the structural route first, so there is less to game. A hint or a retry after a miss caps an item at level 2 and pays a quarter instead of 1. A wrong answer pays nothing but costs nothing beyond tempo; nothing dies. Every answer is three boxes with near-miss distractors, so a child who guesses is right one time in three and the levels know it: an item reaches level 5 only with quick first-try rights on two different days, and a rung passes only when 80% of what was seen is solid. A single question is done after one right, which a guess can earn, so the levels and the ladder, not the done list, are what say a child knows something.

Why "fast in games" stays separate from "known in reading"

The grown-up view in Peabody keeps two columns that other products merge. "Fast in games" is what the practice brain saw: three boxes, a clock, a right tap. "Known in reading" is a different test: a word counts as known only after three right cold reads, with no hint and no boxes, on three different days, with keeper checks at a week, a month and three months. The details are in how Peabody tracks words.

They are kept apart because they are different evidence for different claims. A child who taps said in 1.8 seconds among sad and side has shown recognition under pressure. A child who reads said cold inside a sentence three mornings running has shown something closer to reading. AI writes the stories; code decides the rest, and the code does not let one column vouch for the other. The same idea gives a parent a heat grid of slow and missed times-table facts rather than a single fluency number.

Why emotion detection is off the table

Sensor-free affect detection, reading boredom or frustration from logs alone, is a real research line, and its own results are the reason not to ship it to children.

Baker, D'Mello, Rodrigo and Graesser (International Journal of Human-Computer Studies, 2010) found boredom, not frustration, to be the state most tied to poor learning and the most persistent (PDF). Baker and colleagues (2012) built log-only detectors of concentration, confusion, frustration and boredom that were "better than chance, but not substantially better" (EDM 2012). A 2023 systematic review of 19 studies found typical agreement with human labels (Cohen's kappa) of about 0.2 to 0.4, found that models trained on fewer than roughly 1,500 students were unstable, and found exactly one detector that had ever run in production (arXiv 2310.13711, 2023). Ocumpaugh and colleagues (British Journal of Educational Technology, 2014) found detectors trained on urban and suburban students did not transfer to rural students, or the reverse (PDF.pdf)). The largest study, on 69,174 algebra students, reached correlations between 0.08 and 0.34 with reported feelings (Hutt, Grafsgaard and D'Mello, CHI 2019).

There is also a legal line. The EU AI Act prohibits emotion recognition in education except for medical or safety reasons, under Article 5(1)(f) (Leverhulme Centre, 2024). Peabody is used in the US today, but the principle travels: a product for children should describe behavior ("slow", "missed", "rushing") and never claim to know a feeling.

What a parent should actually get

One plain sentence and one action that takes two minutes. Not a dashboard, not a percentile, not a color-coded grid of forty skills.

Peabody's free grown-up view is built that way: which practice words are known, being learned or missed, and when; which books ran too easy or too hard; a grid of slow and missed times-table facts; and one thing to practice tonight. There are no scores for kids, no leaderboards, no comparison to other children, and everything can be exported or deleted at any time.

What this doesn't show

None of this is a diagnosis. Peabody does not diagnose reading or learning difficulties, and a "slow" or "shaky" label describes taps on a tablet, not a child. Massachusetts' sample family letter for reading screening says the screening "is not a diagnosis or determination of a disability" (Massachusetts DESE, 2024), and Colorado's sequence (screen, extra teaching, monitor progress, and a diagnostic evaluation only if a child is still not progressing) is the right model for what any app may claim (Colorado Department of Education, 2024).

Peabody also cannot hear your child read. It records no audio and scores nothing by ear (a 🎤 button only turns speech into typed words), so it cannot measure fluency or decoding aloud. Every answer is a tap, so a lucky guess can look like knowledge, which is why no single right answer ever counts. And Peabody has not been tested in a trial; the thresholds above are design choices, not validated cutoffs.

What you can do tonight

Open the grown-up view and read the one sentence. Then do the two-minute thing it suggests. If it names a word, ask your child to read it to you inside a sentence, cold. If it names a fact, ask it at the dinner table and count to three in your head. If you have no account, you can still pick a story from the free library one floor below where your child reads and listen for where it gets smooth.

Learn more elsewhere

• J-PAL: teaching at the right level: 13 randomized trials of teaching children at their level, not their grade.
• Duolingo: how we learn how you learn: how one big app predicts when each word is about to be forgotten. A company blog, but it reports real tests.
• NSW CESE: cognitive load theory: the clearest free explainer of why the right help depends on what a learner already knows.

For about seventy more, checked and sorted by who they're for, see where to learn about AI.

Questions parents ask

How does an app know my child is stuck rather than just slow?

Pattern over time. A single slow answer means little. Several misses on the same skill across the last ten or so tries, with no three-in-a-row run of rights, is the definition researchers use for being stuck. Peabody's version is shaky items, a miss that returns three questions later, and a rung that stays put until 80% of what was seen is solid.

Does Peabody time my child?

It measures time on every answer and never demands it. New items have no clock, a slow right answer still counts as right, and speed only raises an item's level after it has been answered right without a hint.

Can an app tell if my child is frustrated?

Not honestly. The best log-only detectors agree with human labels at a kappa of 0.2 to 0.4 and need thousands of labeled students. Peabody describes behavior and never claims a feeling.

Is this a reading assessment?

No. It is a within-Peabody placement based on taps, with no norms, no percentile and no diagnosis. Peabody doesn't measure Lexile, and if you are worried about your child's reading, talk to their teacher.