Curriculum¶

How the eight deck sections, the ten notebooks, the exercises and the interview questions map onto each other — and how to run this as a cohort.

The map¶

graph LR
    subgraph S1["§1 Foundations"]
        N1["NB 01"] --> E1["EX-01..03"] --> Q1["Q1 · Q17"]
    end
    subgraph S2["§2 The eval set"]
        N2["NB 02"] --> E2["EX-04..06"] --> Q2["Q1 · Q16"]
    end
    subgraph S3["§3 System design"]
        N3["NB 03"] --> E3["EX-07..10"] --> Q3["Q4 · Q10 · Q11 · Q12 · Q13"]
    end
    subgraph S4["§4 Retrieval"]
        N4["NB 04"] --> E4["EX-11..15"] --> Q4["Q3 · Q7 · Q8 · Q9"]
    end
    subgraph S5["§5 Context"]
        N5["NB 05"] --> E5["EX-16..18"] --> Q5["Q14 · Q15 · Q16"]
    end
    subgraph S6["§6 Evaluation"]
        N6["NB 06"] --> E6["EX-19..20"] --> Q6["Q2 · Q6 · Q18"]
    end
    subgraph S7["§7 Cost"]
        N7["NB 07"] --> E7["EX-21"] --> Q7["Q5"]
    end
    subgraph S8["§8 Agents"]
        N8["NB 08"] --> E8["EX-22"] --> Q8["Q5"]
    end
    S1 --> S2 --> S3 --> S4 --> S5 --> S6 --> S7 --> S8
    S8 --> CAP["NB 09 · CAP-01<br/>the build brief"]
    CAP --> PF["docs/00-orientation/portfolio.md<br/>the decision record<br/>goes on your CV"]

    classDef sec fill:#FBF1E2,stroke:#E9A83C,color:#101318
    classDef cap fill:#E9F3EE,stroke:#3F8F6E,color:#101318
    class N1,N2,N3,N4,N5,N6,N7,N8 sec
    class CAP,PF cap

Section detail¶

§	Notebook	Core idea	The number it teaches	Exercises	Interview
1	01	The pipeline is a recall budget spent on precision	Recall@N is a ceiling; Recall@k is a delivery	EX-01–03	Q1, Q17
2	02	One record, three evaluations	Full-chain recall, not average evidence recall	EX-04–06	Q1, Q16
3	03	Index-time and query-time are two systems with two SLAs	Storage × and index cost sit next to recall	EX-07–10	Q4, Q10–13
4	04	Staging: cheap-wide then expensive-narrow	ANN recall is a tunable, measured against flat	EX-11–15	Q3, Q7–9
5	05	The context window is the one budget with a hard wall	Marginal full-chain recall per 1k tokens	EX-16–18	Q14–16
6	06	An end-to-end score never says which stage moved	Cohen's κ, and the noise band	EX-19–20	Q2, Q6, Q18
7	07	Cost is a design constraint, not an afterthought	Cost per answered query, to the cent	EX-21	Q5
8	08	The sufficiency check is the whole design	Evidence retention	EX-22	Q5
—	09	Harness first, then one change at a time	The decision record	CAP-01	all

Running it as a cohort¶

Two-day intensive (the format this was built for)¶

	Morning	Afternoon
Day 1	§1–2 · notebooks 01–02 · EX-01, EX-06 in pairs	§3–4 · notebooks 03–04 · EX-07 or EX-12
Day 2	§5–6 · notebooks 05–06 · EX-16, EX-19 (needs pairs)	§7–8 + capstone kickoff · notebooks 07–09

Between the days: post one Q&A discussion and one design review. The overnight thread is where the second day's questions come from.

Eight-week part-time cohort¶

One section per week, plus a capstone fortnight.

Monday: faculty posts the reading assignment as an issue (type: reading, cohort). Students answer the questions as comments — those answers become the seminar agenda.
Wednesday: 90-minute session walking the notebook live, stopping at every measurement.
Thursday–Sunday: the week's exercise, submitted as a branch + exercise issue.
Following Monday: two students present in Show & Tell before the new reading drops.

Weeks 9–10 are CAP-01, presented as a design review.

Facilitation notes¶

Stop at every negative result. The three findings that contradict expectation (§Three results in the README) are where the learning is. A room that only sees confirmations learns to expect confirmations.

Make someone defend a number out loud. "Your full-chain recall went up 4 points — is that a result?" is the most useful question a facilitator can ask, and the answer should be a question back about the noise band.

Run the fault-isolation tree live, on a failure the room picks, before you explain it. The tree makes sense in about ninety seconds when it is producing a verdict on something real, and takes twenty minutes to explain in the abstract.

Do not skip the honesty inventory in notebook 00. Every cohort has someone who will later quote a number from these notebooks in a client meeting, and that table is what stops them.

Prerequisites¶

Required: Python (comfortable reading and modifying), basic probability, the idea of cosine similarity. You do not need transformer internals, GPU experience, or a background in information retrieval.

Useful but not assumed: SQL, having shipped something with an LLM in it, having been on call for anything.

What "done" looks like¶

A student has finished this curriculum when they can, without notes:

Draw the four-plane architecture and say what each plane owns.
Take a bad answer and attribute it to a stage with a trace.
Say what Recall@N bounds and why a reranker cannot exceed it.
Explain why post-filtering an ACL does not satisfy "not influenced by".
Price one grounded answer to the cent and name the three biggest levers.
Say what a noise band is and refuse to call a smaller delta a result.
Produce a decision record that names what they rejected and why.

Six of those seven are about judgement rather than technique. That is deliberate.