The Mastery Engine

10 Minuten · 6 Abschnitte

The part that decides what comes next

In the mastery mode of a Fido day, each child practices on a tablet at their own level. Something has to choose the next exercise, bring back what is fading, and say when a topic is done. That part of the software is the mastery engine. This article explains what it decides, what it leaves to the adults, and how it works, as of November 2026. It is written for parents, educators, funders, and anyone who might help build it.

The Shape of Fido's Software names the parts of the software; this article takes one of them further. Why practice until mastery and review at growing intervals suit children is told in Seven Kinds of Knowledge and the Curriculum Map and in A Day at Fido. This article does not repeat it.

As in the whole series, we say what runs today, what is designed but not built, and what is only proposed.

Three Decisions, and What Stays With the Adults

The mastery engine runs the practice in the mastery mode. It makes three decisions for each child. It picks what the child practices next. It decides when something already learned comes back for review. And it decides when a topic counts as mastered.

It makes these decisions the same way for every child, by rules that anyone can read. A rule says, for example, how many correct answers in a row a topic needs. The people who write the curriculum set these rules for a single topic or a whole branch of topics, as plain data. Changing a rule does not change the program.

Much is deliberately left to the adults. The guide who looks after the mastery mode cares for motivation, finds out why a child is stuck, and keeps the relationship. The engine keeps serving practice; the guide does not reteach the content. Adults also set the curriculum, and they place a newly arrived child by what the child can show, never by age. A guide can record that judgment and undo it later; undoing it leaves alone what the child has earned by practice since.

The child keeps two choices of their own: which course to work on today and, where a sitting offers several, which subject. Inside a course, the engine sets the order; at most it lets the child pick among a few topics it suggests.

One Record, Everything Else Worked Out

Underneath the three decisions lies one idea. Everything a child does is recorded once, as an entry in an event log: this child, this exercise, this answer, right or wrong, at this time on this day. Entries are only ever added, never changed. Every view of the child's progress is worked out from this record. Where a result is kept to save time, working it out again from the log must give exactly the same result.

What does one answer turn into? Step through with +.

The figure follows one correct answer to 7 · 8. The tablet adds it to the log. From the log, the engine works out whether the fact now counts as known, when it should come back, and what the child should do next. On the server, code from the same engine works out the guide's view.

This matters to a parent in four ways.

  • One truth. The child's screen and the guide's table read the same entries, with code from the same engine. When two parts of a program keep their own counts instead, they drift apart without anyone noticing: one says a card is known, the other says it is new.
  • Nothing in a second place. No part of the software keeps a private note of what a child did. What is not in the log counts nowhere.
  • The same answer on tablet and server. The engine is written once and built into both. Given the same entries and the same moment, it gives exactly the same answer.
  • It works offline. The engine runs on the tablet itself. A child can practice a whole morning without a connection. The entries reach the server later, and none is lost or counted twice.

There is a quieter benefit. If we find a mistake in a rule, we fix the rule and work everything out again from the log. No child's history has to be rewritten.

How the Engine Judges That a Child Knows It

The mastery engine never takes one right answer as proof. It judges mastery from several attempts, with a kind of rule chosen for the content.

Several correct answers in a row. For a first skill, such as a letter or one step of a calculation, a topic counts as mastered after a short run of correct answers to different exercises, usually two or three. Answering the same exercise again does not count, and one wrong answer starts the run again.

Remembering over time. For facts and words, the engine estimates how well the child will remember each one, with a published memory model from research on spaced repetition. It plans each review for about the moment the chance of remembering drops to nine in ten. Each successful recall makes the memory steadier, so the gaps grow. With the default settings, a child who answers a new fact correctly each time it comes due, practicing each morning, sees it again after a day, three days, nine days, almost four weeks, about two months, and about five months. No gap grows beyond a year. The fact counts as mastered once its memory is expected to last almost two weeks. That takes correct answers on several separate days; here, the fourth answer, about two weeks in.

Reading enough of a text. For articles and texts, mastery means having spent long enough on enough of the sections. This article is counted that way.

A topic made of parts. A larger topic counts as mastered once enough of its parts are.

Two rules hold for every kind. Mastery earned by practice is never taken back by a later slip; a wrong answer brings the topic back for review sooner instead. And the bars are set as data, per topic or branch of topics. Which single bar Fido should name to parents as "mastered" is still undecided.

How the Engine Picks the Next Step

The curriculum is a knowledge graph, in which every topic knows what it builds on, and a course lays out a path through it. The mastery engine works along that path.

When a child starts, the engine ranks the course's topics, mainly by a fixed order of reasons. A topic whose review is due comes first. Next comes a topic the child left unfinished in the past day, then a topic not yet started, then one in progress. Where two topics are equal, the course's own order decides, never chance. The engine offers up to three topics, and the child picks one.

Inside a topic, the engine picks the next exercise, or it stops. It stops when the topic is mastered, when nothing is left to show, or when the child has had enough for now: after twelve attempts in one sitting, or five misses in a row. A child who is stuck moves on to something else instead of failing again and again.

In a course such as the times tables, the engine composes each sitting as a set of cards, chosen by need. Cards the child had learned and then got wrong come first. Next come cards that stand between the child and the next row of the course. Then come new cards and cards still climbing. A new row opens only once the cards before it reach the course's bar, and until then the sitting deals none of its cards.

One part of the design is not built yet. The engine is meant to check a topic's prerequisites itself and, when a child keeps making the same error, to send them back to the earlier skill that is missing. Today it follows the course's order and does neither. Spotting a missing earlier skill is, for now, the guide's job.

What Guides and Parents See

Guides see their class through views the server works out from the synced event log, with code from the engine the tablets run. For each child and each subject, a guide sees how many topics the child has mastered and how many are in progress. That overview still reads some rules more simply than the tablet does; making it read every rule the same way is still to do.

In a course such as the times tables, the guide's view also points to where help is needed. It lists the rows in which a child has made no progress over their last three sittings, and the facts a child has learned and lost again three times. It shows a wrong answer a child keeps giving, such as 54 for 6 · 8: that is the answer to 6 · 9, so the answer itself names the mix-up. And it lists rows whose review is long overdue. Each entry is a reason for a conversation, not a mark.

Some rules about what is shown hold everywhere. A child sees their own progress and never a classmate's name or place. A guide's view of a group shows how the children are spread out, not a ranking. The scores the engine uses to rank topics are never shown to a child.

Parents are to see their child's progress, worked out from the same log. Today that view exists only in a demonstration, with sample data; The Shape of Fido's Software lists it among the missing parts. The engine's reading is also only one of four signals Fido uses to judge a child's progress, as How We Recognise Progress explains.

What Runs, What Is Proposed, What We Do Not Know

What runs today is what this article has described: the event log, the rules for judging mastery, reviews at growing intervals, the ranking of topics and the composition of a sitting, a guide's placement and its undo, and the guide's views. Once a child has reached a subject's core stages, the engine also works out long-term checks after three months, six months, a year, and then yearly. No screen shows them yet.

Designed, not built, are two parts. The engine is to check prerequisites and send a child back to a missing skill. And each subject is to get a pace line, which tells whether a child is ahead, on pace or behind, judged only by what they still know weeks later, so that time can be shared out fairly. Until the engine works these out, guides apply them by hand.

Only proposed. Two larger ideas are proposals, and neither runs for children. One would plan each child's whole day, with slot, room, and adult, for guides to accept or change each morning. The other is a mathematical model that would turn "mastered" from a rule of thumb into a claim the data can back, with a stated confidence. First calculations show that such certainty for every fact of a short course would take many times the practice the course has, so the claim will have to be narrower.

What we do not know yet. The memory model was fitted on adults reviewing flashcards, and the other numbers in the rules were set by judgment. How well they fit children has not been measured. Nor is it known whether an engine aimed at a sharper goal would help children learn more; as far as we can tell, no one has shown that anywhere.

The series goes on with The App, the Server, and Where a Child's Work Lives and How We Know It Works.

Questions, objections, or keen to help build it? Write to us.

Konzepte

Mastery-EngineEreignisprotokollOffline-BetriebWiederholungWissensnetzCoach-Rolle