Independent assignments

Each chapter now has three distinct activities: an executed worked example in the lesson, a student notebook with an unfinished function and independent acceptance fixtures, and a transfer experiment on new inputs. The 14 student notebooks intentionally contain NotImplementedError; they are not failed book builds or completed solutions.

Read A.11 before running any downloaded notebook, and A.12 before Chapter 2 uses compact syntax. Keep the notebook anywhere inside the cloned repository. A standalone download outside that tree needs GENAI_COURSE_ROOT pointing to the clone. Restart the kernel and run all cells after implementing the function. The published chapter notebooks remain executed reference material.

Chapter	Transfer focus	Download
1	Fixed counts, not a lucky simulation	Student notebook
2	A new design matrix	Student notebook
3	Identity survives reordering	Student notebook
4	Unequal attention weights	Student notebook
5	Detect a deliberately leaky output	Student notebook
6	A stopping token is not grammaticality	Student notebook
7	Optional branch: detect mass between modes	Student notebook
8	Optional branch: an independent score oracle	Student notebook
9	Optional branch: count the update	Student notebook
10	A valid citation can support a false statement	Student notebook
11	Join by question, then subtract	Student notebook
12	Reject a plausible invented finish	Student notebook
13	Compute an interaction after matching tasks	Student notebook
14	Honor every requested setting	Student notebook

The acceptance fixtures use hand-computable targets and negative cases, independently of the worked functions. They are learning checks, not a hidden test set. A function can pass a few fixtures and still be wrong elsewhere: add your own counterexample. In particular, a constant zero gradient, uniform attention that ignores logits, and a constant causal-test pass are rejected.

What counts as completion

A complete submission includes the implemented function, executed fixtures and a deliberately wrong control, all rows of the prescribed transfer experiment, and an interpretation naming the estimand and the limit of the evidence. Use the notebook’s 3+2+3+2 rubric. Numerical agreement is distinct from support for a scientific claim; an unexpected comparison is acceptable when correctly recorded and interpreted.

For Chapters 10–14 use the common evaluation record. A missing review stays null. A valid JSON response with an unsupported claim is a failure. A refused unsupported capstone request can be correct behavior, but only if its intent was preserved. Chapters 7–9 are optional: their small independent numerical experiments are sufficient; training a large image or language model is not required.