Most classrooms assume everyone starts in the same place and moves at the same speed. If you got a 72 on the unit test, you moved on with a 28% hole, and the next unit was built on top of it.

In software, building the next layer on a broken one is how you get paged at 3 a.m. When I learn a new field, I want what I'd want in a system: know the dependencies, know the current state, fix what's broken first.

That led me to two old ideas from education research, knowledge space theory and mastery learning, and to a simple AI workflow that puts them together. It runs from a folder and a set of written instructions. No app required.

I started using it to dust off my Chinese. It has since spread into other topics: taxes, macro finance, medicine, business acquisition, and more.

What is knowledge space theory?

Knowledge space theory (KST) was introduced by Jean-Paul Doignon and Jean-Claude Falmagne in a 1985 paper, "Spaces for the assessment of knowledge". The idea is easy to state.

Treat a subject as a finite set of concepts. Some depend on others: you can't do much with derivatives without limits and functions. Your knowledge state is the exact set of concepts you've got right now. Not a score, a set.

Because of prerequisites, not every combination is realistic. The collection of feasible states is the knowledge structure. In the cleanest version, called a learning space, any state can be reached by learning one concept at a time.

The part I find most useful is the fringe:

• The outer fringe of your state is the set of concepts you haven't learned yet but whose prerequisites you already have. It's what you're ready to learn next.
• The inner fringe is what you most recently added: the newest, least secure pieces.

Instead of "you're at 62% in algebra," KST says "here is exactly what you know, and here are the five things you're ready for."

ALEKS: knowledge space theory in production

The best-known application is ALEKS (Assessment and LEarning in Knowledge Spaces), built on this theory. According to ALEKS, Algebra 1 is modeled as roughly 350 basic concepts, which produces millions of feasible knowledge states. Even so, an adaptive assessment can pin down a student's state in about 25 to 30 questions, because each answer rules out large parts of the structure.

ALEKS uses this vocabulary directly. Its teacher materials describe the "ready to learn" list as the outer fringe of the student's state, and the most recently learned topics as the inner fringe, which are "considered the least secure and most likely to need reinforcement."

Building the map has always been the expensive part. Knowledge structures are usually built by querying experts with long series of "does X require Y?" questions, or by mining student data. That cost is a big reason the approach mostly lives inside a few commercial products.

What is mastery learning?

Mastery learning is simpler: don't move on until you've actually got it.

Benjamin Bloom laid it out in "Learning for Mastery" (1968). His argument was that most students, "perhaps over 90 percent," can master what teachers have to teach, and he treated aptitude as the amount of time a learner needs to reach mastery. Around the same time, Fred Keller published "Good-bye, Teacher..." (1968), describing what became the Personalized System of Instruction, or the Keller Plan. Its core features included self-pacing, a "unit-perfection requirement" before advancing, and student proctors who made repeated testing, immediate scoring, and tutoring possible.

Both run the same loop: break material into units, check, correct, re-check, advance.

Bloom's 2 sigma problem

In 1984 Bloom published "The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring". Drawing on dissertation studies by his students Joanne Anania and Arthur Burke at the University of Chicago, he reported that the average student taught one-to-one by a tutor using mastery learning scored about two standard deviations above the average student in a conventional class. In his words, the average tutored student was "above 98% of the students in the control class." Students in group mastery learning landed about one standard deviation up.

Tutoring showed most students could learn at a much higher level, but it was "too costly for most societies to bear on a large scale." The problem was finding group methods that come close.

What the evidence actually says

The 2 sigma number gets repeated far more confidently than the research supports.

• The tutoring effect is real but usually smaller. Kurt VanLehn's 2011 review, "The Relative Effectiveness of Human Tutoring, Intelligent Tutoring Systems, and Other Tutoring Systems", found an average effect of about d = 0.79 for human tutoring and d = 0.76 for step-based intelligent tutoring systems. Large, but well short of 2.0. VanLehn also noted Bloom's tutored students were held to a higher mastery bar (90% vs. 80%).
• Mastery learning works, with caveats about measurement. A 1990 meta-analysis by Kulik, Kulik, and Bangert-Drowns found positive average effects of roughly half a standard deviation across mastery programs, including Keller's PSI. But Robert Slavin's "Mastery Learning Reconsidered" (1987) looked at a narrower set of studies he considered higher quality and found effects near zero on standardized tests, with gains mostly showing up on tests written by the experimenters around the same objectives. He also noted mastery learning usually takes extra time, which complicates comparisons.
• The best summary I've found is José Luis Ricón's review of the 2 sigma literature: tutoring probably doesn't hit 2 sigma on average, mastery learning probably isn't a full 1 sigma, but both work, and high-quality tutoring, human or software, can reach very large effects.

My takeaway: individual attention, frequent checks, and refusing to build on gaps all help, most when targets are clear.

Why knowledge space theory and mastery learning fit together

Knowledge space theory answers what to learn next. It gives you a map, places you on it, and tells you which concepts are within reach.

Mastery learning answers when to move on. It says the outer fringe only expands when the concepts you just learned are actually solid.

Without the map, mastery learning is a fixed sequence, even for units you already know. Without mastery, the map is just a fancier syllabus.

Put together, you get a loop: find your state, pick something on the outer fringe, work until it's mastered, update your state, repeat. When you fail, the map says which prerequisite to revisit instead of just "study more."

Why AI makes this practical now

For decades, both halves were expensive. Maps took experts and data. Tutoring took a person. Large language models change the cost of each piece:

• Generating a domain map. A model can draft a concept graph for almost any field in minutes, playing the "expert" KST researchers used to query.
• Adaptive placement. It can run a placement test in the spirit of ALEKS, choosing each question based on the last answer.
• Instant feedback. It can read your actual work and tell you where your reasoning broke.

Early evidence is encouraging. A 2025 randomized trial in a Harvard physics course, Kestin et al. in Scientific Reports, found students using a carefully prompted AI tutor learned more in less time than in an active-learning class covering the same material. It was one course, two lessons, and custom post-tests, so I treat it as promising, not settled.

The design matters a lot. In a field experiment with nearly a thousand high school math students, Bastani et al. (PNAS, 2025) found that students with unrestricted GPT-4 access did better during practice but 17% worse on the exam once access was removed. A guardrailed version that gave hints instead of answers largely avoided that harm. If the AI does the thinking, you don't learn.

The caveats

• Models can be confidently wrong. They're trained and graded in ways that reward guessing over saying "I don't know". An explanation can sound perfect and be wrong.
• The map needs checking. A generated graph is a draft. Compare it against a good textbook or syllabus. Missing or reversed prerequisites send you the wrong way.
• Mastery today isn't retention next month. The research on effective learning techniques rates practice testing and spaced (distributed) practice as the highest-utility strategies. A mastery loop without spaced review still leaks.

How my AI mastery learning workflow works

The whole thing is a folder plus written instructions. I run it in Claude Cowork, but any AI tool that can read and update files in a folder can run the same pattern.

What goes in the folder

My setup is deliberately small: two core files plus the instructions.

• Instructions. How to build the map, run placement, run sessions, and what counts as mastery.
• Concept map. Each concept with its prerequisites, a few ways to check it, and its current status: mastered, shaky, or not yet learned. The map doubles as my knowledge state, so the frontier is visible right on it.
• Log file. A running record of each session: what was tested, what broke, what was taught, what changed, and any misconceptions worth coming back to.

Keeping state in files means the AI doesn't have to remember you; it reads where you left off.

Stage 1: Build the concept map

The instructions tell the AI to map the field as concepts with explicit prerequisites, each small enough to check with one or two questions. Too coarse ("calculus") and you can't place anyone. Too fine and the map is unusable.

It is worth checking the map against a trusted source for missing foundations and anything out of order.

Stage 2: Run the placement test

The instructions tell the AI to start in the middle of the map, move up on correct answers and down on misses, ask me to explain my reasoning, and stop once it can mark each area known, unknown, or shaky.

The output is the map with every node marked. What matters most is where the edge nodes are: concepts whose prerequisites are solid but which I haven't mastered yet. That edge is the outer fringe from knowledge space theory, and it's where every session starts. The test isn't there to grade me. It's there to find the frontier.

Stage 3: Run the feedback loop

Each session follows the same loop:

1. Apply. The AI gives me a problem or task that uses a concept from my outer fringe. Real use, not "define X."
2. Expose. It looks at my work and finds where my understanding is wrong or thin. Not just "incorrect," but which assumption broke.
3. Teach. It traces the gap to a prerequisite or adjacent concept on the map, teaches that briefly, and has me try again.
4. Advance. Once I can do it cleanly, more than once and in a slightly different form, it marks the concept mastered, updates the map, and picks the next concept on the fringe.

The instructions also say what not to do: don't give the answer before I've tried, don't mark something mastered from one correct answer, and log every misconception for later review.

The loop also finds holes I didn't know I had. It caught gaps in my Spanish grammar, and Spanish is a second language I'm really strong in. That's the point: everyday fluency hides gaps, because you route around what you don't know without noticing. A test that probes each node doesn't let you.

Try it yourself: learn anything with AI

The setup is small.

1. Make a folder for the field you want to learn.
2. Write the instructions. Describe the three stages in plain language. Define mastery concretely, e.g. "solve two new problems without hints and explain why."
3. Ask the AI to build the concept map and save it to a file. Then check it against a textbook or syllabus.
4. Take the placement test. Be honest. Say "I don't know" instead of guessing.
5. Run short sessions through apply, expose, teach, advance. Have the AI update the map and the log at the end of each one.
6. Add review. Start each session by re-testing a few earlier concepts, especially recent ones and logged misconceptions.

The main rule: you do the work. The AI picks the next problem, finds the flaw, and points you at the missing piece. If it starts solving things for you, tighten the instructions.

What's next

This started as a tool for myself, and I think it's a fresh take on education: placement instead of a fixed starting line, a map instead of a syllabus, and a tutor that won't let you build on gaps. That anyone can run it with a folder and some instructions is the part I like most.

I may package it as an app at some point so people don't have to set it up by hand. No name or timeline yet. Until then, the folder version works. If you try it, I'd like to hear what your placement test found.