This guide is for a teacher deciding what to teach from Trace and how much of it fits the time available. It covers the whole grades 11-12 course rather than any one chapter. Each chapter that has a teacher pack also has its own overview, prerequisites, and pacing notes under Resources.
What this course is#
Students learn how AI systems work by reading, experimenting in supplied labs, and building. At grades 11-12 they compare evidence, inspect the tradeoffs in training and serving a model, and design bounded uses of one, and the same four-step habit runs through every chapter: identify the representation, find the decision or computation, check the evidence that bears on it, and keep the claim inside the conditions that were tested.
A student who finishes the front half can take apart an ordinary product and say what goes in, what comes out, who chose its objective, how it will probably break, and what number would settle the argument. After the language-model chapters they can trace one request from text to answer and say what that answer does and does not establish.
The course is not a coding course, although an optional Python code track is designed for several chapters and its packages are waiting on app support. It is also not a tutorial for any product. Real companies, models, and platforms appear as dated examples, never as things to learn to operate.
How a chapter is put together#
A chapter holds units, a unit holds packages, and a package holds the activities a student actually opens. One package is one square on the chapter grid, covering one or two closely related topics through one to five articles, one lab, and one quiz. A unit is a tested arc of two to four packages and closes with a unit test. Package and unit counts vary by chapter, because depth is decided per topic rather than filled to a quota.
Every chapter is meant to open with an introduction unit, u0-introduction:
one or two plain-language articles on what the topic is, what the chapter is
for, and where the student already meets it, with no quiz and no unit test
and no bearing on mastery. It may carry a demo lab that is pure play with the
finished thing the chapter will explain. Chapter 10 has this unit today; the
other live chapters pick it up as each comes up for republication.
What mastery and coins mean, and what they do not#
A package counts as proficient when the student has saved a lab run and answered every question on the package quiz correctly. Either half alone is incomplete, so a student cannot finish a package by multiple choice. A unit counts as mastered when the student has passed its unit test at 80 percent and holds every package in the unit proficient. The unit test draws 25 questions per attempt from the unit's bank.
Coins are progress records. The defaults are 10 for a proficient package, 50 for a mastered unit test, and 100 as a chapter bonus, and students spend them in the Shop. Trace does not assign a grade, and neither coins nor proficiency is a score.
A package quiz is five fixed multiple-choice questions, and a student can retake it until all five are correct. The same five questions come back on the retake, so a student can pass by remembering which option was right last time. An all-correct quiz is evidence that the student worked through the package, and it measures less than it looks like it measures. When you want your own check on what a class knows, use the assessment bank in the chapter's pack under Resources. It is a separate pool from the student quizzes and unit tests, tagged by concept, difficulty, and intended use, and you can assemble a worksheet or a test from it.
The chapter map#
Twelve chapters are live today. Counts below are the units and packages in the published chapter, and the last column names the chapters whose material each one builds on by name rather than reteaching.
| # | Chapter | What it teaches | Units | Packages | Builds on |
|---|---|---|---|---|---|
| 1 | AI Around Us | Recognizing learned systems in ordinary products: inputs and outputs, goals and incentives, failure patterns, evidence, generated media, and a written judgment | 7 | 18 | The opener; assumes nothing |
| 2 | Data, Patterns, and Labels | How decisions turn the world into a dataset, then labels and proxies, who is in the file, what cleaning decides, and provenance | 5 | 15 | 1 |
| 3 | Classification | How a classifier decides with a score and a threshold, and how to judge one: confusion matrix, base rates, calibration, error costs, per-group rates | 7 | 13 | 1, 2 |
| 4 | Regression | Predicting a number: least squares by hand, residuals, error measured against a baseline, validation, and drift | 8 | 16 | 1, 2, 3 |
| 5 | Unsupervised Learning | Finding groups with no answer key: distance, k-means, hierarchical clustering, and judging whether a grouping is real | 6 | 15 | 1, 2, 3, 4 |
| 6 | Responsible AI | Fairness, privacy, transparency, accountability, and governance, worked through documented cases | 6 | 14 | 1, 2, 3 |
| 7 | Neural Networks | A neuron computed by hand, the forward pass, loss and gradient descent, backpropagation, and reading a training run | 7 | 12 | 1, 2, 3, 4 |
| 8 | Reinforcement Learning | Learning from reward: bandits, grid worlds, discounted returns, Q-learning, and reward design failures | 6 | 15 | 1, 2, 3, 4, 5, 7 |
| 9 | Neural Network Architectures | Convolution for grids, recurrence and the LSTM for sequences, encoder-decoder with attention, autoencoders and GANs | 6 | 14 | 7, plus 1-5 |
| 10 | Modern LLMs and AI Agents | One request traced from tokens through attention and decoding, then training, serving, reasoning modes, chat apps, multimodality, and the electricity behind it | 10 | 35 | 1, 2, 7, 9 |
| 14 | Prompting and Reliable Outputs | A prompt as an engineering artifact: specified, tested against a suite, versioned, with structured output and guardrails priced | 6 | 18 | 1-10, every callback retaught in place |
| 15 | How AI Agents Work | The agent loop and the transcript, the tool round trip, context as a budget, planning, verification, stopping, and permissions | 6 | 13 | The front half only; unit 1 retaught for a class that skipped 10 |
Chapter 10 is the only chapter with a teacher pack today, and that pack is still being finished, with one of its nine slide decks built. It is also the only chapter with a planned lesson count, 45 lessons across its nine taught units. For the others, plan from the unit and package counts above until their packs arrive.
Ten further chapters are planned and not yet written: Generative Media, Training Language Models, Inference and Model Optimization, Enterprise AI Systems, Evaluating AI Systems, the AI Systems Capstone, and an extension strand on chips and fabs, data centers and grids, materials and labor, and AI in attention and relationships. Their numbering above 10 is still moving, so plan around the live chapters and treat the rest as coming.
Fitting the course to your class#
These are recommendations drawn from the dependency column above, not rules. Chapters name earlier material rather than reteaching it, so skipping a listed prerequisite means patching it yourself. Each live chapter's pack, as it is written, says what to patch and how.
For one semester, chapters 1 through 5 cover the classical machine-learning core and chapter 6 closes it on judgment and governance. If that runs long, chapter 6 needs only chapters 1, 2, and 3, so chapter 5 can come out without breaking anything. For a full year, run 1 through 10 in order and add 14 and 15 if the calendar holds. For a short elective of six to eight weeks, chapter 1 followed by chapter 15 works, because chapter 15's first unit reteaches what a language model is for a class that never read chapter 10. For a single unit inside another course, chapter 1's unit 1, "What counts as AI", needs no preparation, and chapter 15's unit 1, "The frame: what the model receives", stands on its own for a class that wants agents specifically.
The natural stopping points are after chapter 5, where the classical machine learning is complete; after chapter 6, which closes on judgment and governance; and after chapter 9, before the language-model chapters begin.
The two ways a class uses Trace#
In a platform block students self-pace through the assigned articles, labs, and quizzes while you circulate, help the students who are stuck, and decide when to stop the whole class for a debrief. Your job is orchestration, and the chapter pack gives you the regroup points, the common stuck points, and a fast-finisher assignment. In a taught lesson you teach from the chapter's slide deck with a worksheet and an exit ticket, and the platform holds practice rather than instruction.
Chapter packs are organized by planned lesson, meaning one suggested 50-minute period. Packages and units vary deliberately in depth and do not map one to one onto periods, so lesson boundaries are editorial recommendations. The lesson timeline marks each lesson CORE or OPTIONAL and marks which lessons expect a platform block.
Math and reading the course assumes#
Students need arithmetic they can do on paper: addition, subtraction, multiplication, division, percentages, and unit conversion. There is no calculus anywhere. Chapter 3 counts natural frequencies from a table, chapter 4 fits a least-squares line by hand, chapter 5 computes distances, chapter 7 computes a neuron and walks gradient descent numerically with slope as the only formalism, and chapter 10 uses simple vectors throughout and watts and megawatt-hours in its last unit.
Articles run roughly 800 to 1,500 words at this band. No term is used before it is defined, including in passing, so a student who reads in order does not need a reference open. The heaviest arithmetic sits in the chapters that compute rather than argue, which are 4, 7, and 8, and chapter 10's pacing notes flag its units 4 and 6 as the longest at six lessons each. Patch a numerical slip as a numerical slip rather than as a conceptual failure.
The honesty bar, and answering "is this true?"#
The content describes what models do rather than what they understand. An article will say a model produces tokens from a numerical context, and it will never say the model knows, believes, or intends anything, which is the line worth holding when a class discussion starts to drift.
Every real fact carries its date. Perishable numbers such as prices, specs, and benchmark scores appear as dated examples and are never the thing being taught or tested. Statistics are real and sourced, with no invented numbers and no recurring fictional characters, and research claims travel with their venue and peer-review status, so a preprint is labeled a preprint and a single unreplicated study is labeled as one. Where the field is still arguing, both sides travel together.
When a student asks whether something is true, the answer is on the page. Read the date, the venue, and the status attached to the claim, and say what that evidence supports and what it leaves open. Completion is never offered as proof that a student has mastered anything general, which is worth saying out loud to a class that reads coins as a ranking.
Safety and privacy in class#
No live products, accounts, credentials, or student data belong in this course. The supplied labs are the hands-on surface, and readings use fictional, non-personal scenarios when they need a context. Keep student work inside the supplied fixtures and the systems you have authorized.
Trace records that work happened: finished activities, saved lab runs, quiz
and unit-test answers, worksheet submissions, and the AI usage and
conversations inside the platform. There is no score column and no grade
export. Trace does not use student work, prompts, or answers to train or
improve any model, and every model request tells the provider not to retain
it. The full policy is at /legal/privacy.
Before the first day#
- Have a Trace account with the teacher role, either from
/sign-upor from your school's invitation. - Create your class and give it a name you can tell apart later.
- Decide how students get in, and either print the join-code sheet from the Students section or paste your roster into Invite by email.
- Assign the first lesson. Chapter 1, unit 1, "What counts as AI", needs no preparation.
- Check the devices. Chromebooks and touch screens work; very old browsers do not.
- Read the chapter overview and lesson timeline for your first unit under Resources, where a pack exists.
- Print the unit's substitute kit and leave it where a substitute will find it.
The platform guide covers each of these steps with the exact buttons. Start with "Getting started for teachers".