Appendix E — Course Syllabus and Week Map
The book is deliberately larger than one semester. This appendix answers “which chapters do I actually have to read, and in what order?”
Thirty-nine chapters do not fit in fourteen weeks, and pretending otherwise produces a course that covers everything badly. What follows assigns about two chapters a week, names what gets cut, and says plainly that the cut material is reference rather than an oversight.
E.1 The fourteen-week course
| wk | chapters | in class | due |
|---|---|---|---|
| 1 | Chapter 1, Chapter 2 | Tokenize Turkish and English side by side; count what breaks | — |
| 2 | Chapter 3, Chapter 4 | Run the bake-off; argue about the baseline | Lab 1 |
| 3 | Chapter 5; reading: Chapter 22’s sizing arithmetic and scorer validation | Annotate 50 items in pairs; measure disagreement | — |
| 4 | Chapter 6, Chapter 7 | Attention by hand on one sentence | Lab 2 |
| 5 | Chapter 8, Chapter 9 | Fertility and total tokens on your own text; price the total tokens from a real API price list, input and output separately | Capstone proposal |
| 6 | Chapter 12, Chapter 13 | Fine-tune one encoder; read the errors | Lab 3 |
| 7 | Chapter 16 | The control-measurement workshop, with twenty minutes on validating a scorer | — |
| 8 | Chapter 17, Chapter 18 | Structured output, and making it fail | Capstone eval set |
| 9 | Chapter 19, Chapter 20 | Build retrieval; measure it alone | Lab 4 |
| 10 | Chapter 22 | Audit a judge; size an eval set | — |
| 11 | Chapter 23, Chapter 24 | Build the smallest agent that works | Mid-point check-in |
| 12 | Chapter 25, Chapter 30 | Trajectory versus outcome scoring | Lab 5 |
| 13 | Chapter 36, Chapter 37 | Attack each other’s agents | — |
| 14 | Chapter 39 | Demos | Capstone due |
Three deliberate choices in that table, each of which someone will disagree with.
Week 7 is spent on Turkish alone. For a course taught at a Turkish university it is the highest-value week in the semester, and Section 16.5’s pattern — standard tools silently wrong outside English — is a lesson that transfers to every other topic.
Evaluation arrives at week 10, before agents. Chapter 22 has to precede Chapter 30 or students build agents they cannot assess.
That leaves one ordering problem to handle rather than ignore: the capstone’s eval-set deadline is week 8, and it asks for a validated scorer and a stated minimum detectable difference — both of which live in week 10’s chapter. Two subsections have to move earlier. Assign Chapter 22’s sizing arithmetic and its scorer-validation argument as reading in week 3, alongside Chapter 5, where measuring disagreement already makes the case; and spend twenty minutes of the week-7 workshop on validating a scorer, which is the same control-measurement discipline that week is about. Week 10 then deepens what students have already had to use, which is the better order anyway.
Week 13 is adversarial and hands-on. Reading Chapter 36 makes the argument; a planted instruction that works on the first try makes it concrete, and that is worth an hour of class.
Set it up so nobody attacks anything they do not own. The exercise runs on a shared class harness: a sandboxed agent with a throwaway data store, provided by the instructor, that every student can attack and that resets between attempts. Attacking a classmate’s system is fine only with that classmate’s written consent and only in that sandbox — Chapter 36’s own condition, applied to the classroom. Nothing in this course licenses testing against a system you were not given permission to test.
E.2 What gets cut, and why that is fine
Fifteen chapters are not on the list — twenty-four assigned, fifteen cut, thirty-nine in total. They are not lesser material, and several are the most immediately useful chapters in the book for someone already working. They are cut because a semester is short.
| cut | chapters | who should read it anyway |
|---|---|---|
| Most of Part VI | Chapter 31, Chapter 32, Chapter 33, Chapter 34, Chapter 35, Chapter 38 | Anyone who will operate a system rather than build a prototype. This is the usual casualty and the most reference-like part of the book. |
| Advanced agents | Chapter 26, Chapter 27, Chapter 28, Chapter 29 | Students whose capstone needs them; read on demand |
| Adjacent tasks | Chapter 11, Chapter 14, Chapter 15, Chapter 21 | Read the one your problem is |
| Depth | Chapter 10 | After the course, when the rest has settled |
Part VI being the casualty is worth saying plainly rather than hiding: a course can teach you to build something that works once, and Part VI is about the difference between that and something that keeps working. Students who go straight into industry should read it first, in that order.
E.3 Three shorter paths
(a) The practitioner who only wants agents. Start at Chapter 23 and read Part V, then Chapter 30, Chapter 36 and Chapter 37. Detour back to Chapter 17 and Chapter 18 when structured output stops behaving, and to Chapter 20 when retrieval does. Roughly ten chapters.
(b) The statistician who wants the modelling story. Parts I–III in order — Chapter 1 through Chapter 16 — then jump to Chapter 22, which is the chapter this book was most obviously written for a statistician. Then Chapter 30 and Chapter 35 for how the same reasoning survives contact with production.
(c) The reader who wants Turkish NLP specifically. Chapter 2 → Chapter 4 → Chapter 16, then the Turkish measurements embedded elsewhere: Chapter 13’s contamination finding, Chapter 14’s ROUGE tokenizer bug, and Chapter 15’s metric asymmetry. Appendix F is the catalogue.
E.4 Assessment and course policies
The table above is a reading map; this is the part that makes it a syllabus. Weights are a starting point for whoever teaches from this, not a rule — but having no weights is the failure mode, because students allocate effort to what is graded and a course that will not say what that is has already chosen for them.
| component | weight | when |
|---|---|---|
| Five labs, submitted as notebooks | 25% — mean of the best four | weeks 2, 4, 6, 9, 12 |
| Capstone proposal | pass/fail gate | week 5 |
| Capstone eval set | 5% | week 8 |
| Mid-point go/no-go review | 5% | week 11 |
| Participation in the week-13 adversarial session | 10% | week 13 |
| Capstone repository, report and demo | 55% | week 14 |
There is no written exam, and the reason is the argument of the book. A number you cannot defend is worthless, and a three-hour exam cannot test whether you can defend one. What replaces it is the ten-minute oral check inside the demo, where every team member answers questions about any part of the submission — which is a harder examination of understanding than a paper is, and a fairer one for students who build well and write slowly.
Labs are due at the end of the week they belong to, and the lowest score is dropped: the lab category is the mean of the best four of five, so each of those four carries 6.25% of the final grade. One bad week should not follow a student through a semester.
The three capstone gates cannot be made up. They exist to stop a project failing at week 13, and a late gate has already failed at that job. Late policy on the final submission is the instructor’s to set — announce it in week 1 rather than in week 14.
Week 13 is the one session we run synchronously, so that students can attack each other’s agents and inspect and discuss the failures together. A shared harness with snapshot agents could run it asynchronously, at the cost of that discussion. Everything else can be caught up from the book.
Academic integrity, including LLM use, is governed by Section D.2 and applies to the labs as well as to the capstone. Accessibility arrangements go through the university’s disability support unit and never require disclosing a diagnosis to the instructor.
E.5 Prerequisites
Concretely, what you should be able to do in week 1 — not what you should have taken.
Required. Read and modify a Python script of a hundred lines. Work with lists, dictionaries and files. Read a table of numbers and say what it means. Set up an environment, install packages and read an error message far enough to act on it — the pip install that fails is a rite of passage rather than a prerequisite, but the not-panicking is the skill.
Assumed but refreshable. Vectors and dot products, conditional probability, and what “minimising a loss” means. Appendix B covers each in two lines with a pointer to where the book uses it; read it in week 1 rather than when you are stuck.
Explicitly not required. Deep-learning coursework, GPU access, calculus beyond derivatives-as-slopes, or prior NLP. The measurements in this book were all run on one laptop, and Appendix A is what you need to reproduce them.
Not negotiable, and not technical. Willingness to report a number that disagrees with what you hoped. The book asks for that on nearly every page, and the capstone rubric in Appendix D grades it.