Appendix E: Reading map

Goal. Pair the reference books to the chapters they serve, and suggest an order to read them in relative to the book's five authoring waves.

I lean on eight core reference books for the theory spine, plus four practical references for the grounding chapters (Part VIII) and the data pipeline (3.9). None is quoted at length anywhere in the text; instead each chapter carries a read-along pointer by short key, and this appendix is the master index behind those pointers. Local copies live in references/ at the repo root (see that directory's README.md); they are not distributed with the book.

The eight books

Key	Book	Primary parts served
[S&B]	Sutton & Barto, Reinforcement Learning: An Introduction (2e)	Part V
[RLHF]	Lambert, RLHF: LLM Alignment and Post-Training	Parts IV–VI (esp. ch. 5–8)
[BRM]	Raschka, Build a Reasoning Model (From Scratch)	Parts III, V, VI, VII (ch. 3–8; App. C Qwen3 source, App. F evaluation)
[BLLM]	Raschka, Build a Large Language Model (From Scratch)	Part I
[MADL]	Chaudhuri, Math and Architectures of Deep Learning	Part I math spine (ch. 2–9)
[GAIA]	Hurbans, Grokking AI Algorithms (2e)	Pre-reading only; RL intuition (ch. 10)
[CAI]	Ness, Causal AI	Part IV
[AIE]	Huyen, AI Engineering	Breadth complement: Parts I–III, VI, and 7.5 (ch. 2–4, 7–10)

Three of these earn special notes. [GAIA] is deliberately light-duty: it overlaps the other texts at lower depth, so it serves as optional pre-reading for RL intuition, never as a chapter dependency. [CAI] earns a full part of its own (Part IV) because judge-model and benchmark-comparison pitfalls are confounding problems, and framing evaluation claims causally is a committee-grade differentiator. [AIE] is the production-engineering complement: it covers the same territory as Parts II, III, and VI at survey depth but from the builder's seat (evaluation methodology, AI-as-judge, finetuning, dataset engineering, inference optimization), so I cite it where a chapter benefits from the systems-level framing rather than for derivations.

Practical and framework references

These hands-on references back the grounding chapters (Part VIII), the data pipeline (3.9), the post-training landscape (6.1), and the agentic-RL product chapter (11.1). They are framework-, tool-, and adaptation-level rather than theory, so I cite them for how to build a thing, not for why it is true.

Key	Book	Primary chapters served
[RAG]	Rothman, RAG-Driven Generative AI	8.1 (naive/advanced/modular RAG, retrieval metrics, evaluation)
[LC]	Auffarth & Kuligin, Generative AI with LangChain (2e)	8.1 (RAG systems), 8.2 (tools, agents, MCP in ch. 9), 11.1 (ReAct, LangGraph, memory)
[GADP]	Lakshmanan & Hapke, Generative AI Design Patterns	8.1 (RAG patterns), 8.2 (agentic AI, MCP in ch. 7), 8.3 (evaluation, ch. 5–6)
[DPA]	Harenslak et al., Data Pipelines with Apache Airflow (2e)	3.9 (DAG anatomy, scheduling, asset-aware scheduling)
[AGENTS]	Infante, AI Agents and Applications	11.1 (LangGraph tool-agents ch. 5/11, multi-agent ch. 12, memory/guardrails ch. 14), 8.2 (MCP servers, ch. 13)
[CUST]	Bahree & Tok, LLM Customization and Fine-Tuning	6.1 (the adaptation continuum, when not to adapt), Part VI
[PDL]	Dürr, Sick & Murina, Probabilistic Deep Learning	Part I (probabilistic view of the LM objective), 3.7 (uncertainty; concepts, not its Keras/TFP code)

[RAG] is the dedicated retrieval text and the closest companion to 8.1: its naive/advanced/modular framing, retrieval metrics, and cosine-similarity evaluation are exactly the moves 8.1 narrows to a single legible pipeline. [LC] and [GADP] are the tool-and-agent references for 8.2: both cover the Model Context Protocol directly ([LC] ch. 9, [GADP] ch. 7) alongside the tool-use and agentic-RAG patterns, so read them for the ecosystem of options and read 8.2 for why a physics oracle behind a typed tool is the version worth grading. [DPA] is the Airflow book behind 3.9; its asset-aware scheduling chapter (ch. 4) is the same data-aware pattern the pipeline uses to rebuild the task asset when a snapshot updates, and its incremental-and-backfill chapter (ch. 3) is the "why an orchestrator, not cron" argument in concrete form.

[AGENTS] and [CUST] back the product-facing and post-training chapters. [AGENTS] is the agent-framework companion for 11.1: its LangGraph tool-agent chapters (ch. 5, ch. 11), multi-agent chapter (ch. 12), MCP-server chapter (ch. 13, which also backs 8.2), and productionization chapter (ch. 14, memory and guardrails) are the scaffolded-agent world that agentic RL learns to drive from inside. [CUST] frames 6.1: its adaptation continuum and its "when not to adapt" discipline are the wide-angle version of this book's post-training landscape, and it works on the same Qwen3 instruct family this book serves. [PDL] is a lighter, concepts-only reference (it is Keras/TensorFlow-Probability, so I cite it for the probabilistic view of the LM objective in Part I and for uncertainty framing in 3.7, not for its code). Two books in this batch I do not cite: Math and Architectures of Deep Learning is already the Part I math spine above ([MADL]), and Evolutionary Deep Learning sits outside this book's scope, since population-based neuroevolution is a different paradigm from single-GPU RLVR and is impractical on one 16GB card.

Suggested order

The books are not read cover-to-cover front-to-back; they are read alongside the parts they serve, in roughly the authoring-wave order (spec §8).

Before anything (optional): skim [GAIA] ch. 10 for gentle RL intuition. Nothing downstream depends on it.
With Part I (the theory spine): read [MADL] ch. 2–9 as the math backbone and [BLLM] ch. 2–5 as the from-scratch build companion. These two run in parallel with Part I chapter for chapter.
With Parts II–III (inference and evals): [BRM] ch. 1–3 and its App. C (Qwen3 source) back the model-anatomy and eval-framing chapters.
With Part IV (causal): [CAI] Parts 1–3, in order, are the spine.
With Part V (RL foundations): [S&B] is the backbone (ch. 2–4 and ch. 13 carry the most weight), with [BRM] ch. 6–7 and [RLHF] ch. 6 arriving as the material turns to LLM-specific policy optimization.
With Part VI (post-training): [RLHF] ch. 1–8 is the throughline; [BRM] ch. 4–5 and ch. 8 cover reasoning-time scaling and distillation.
Anytime, in parallel (optional): [AIE] reads well cover to cover as the production-context companion; its foundation-model survey (ch. 2) pairs with 1.7 and 6.1, evaluation chapters (ch. 3–4) with Part III, finetuning (ch. 7) with Part VI, dataset engineering (ch. 8) with 3.8/7.5, inference optimization (ch. 9) with Part II, and its architecture chapter (ch. 10) with 3.12.
With the pipeline and grounding chapters (3.9, Part VIII): [DPA] ch. 2–4 alongside 3.9 (its asset-aware scheduling chapter is the closest match to the Airflow-3 pattern used there); [RAG] and [LC]'s RAG chapters alongside 8.1; and [LC] ch. 9 with [GADP] ch. 7 alongside 8.2, since both cover MCP directly. These are consulted as build references when a chapter's lab is in front of you, not read end to end.

Per-chapter pairings

Every entry below mirrors the read-along admonition in the named chapter. Chapters with no reference pairing are hands-on or self-contained (parts of Part II and Part X lean on tooling and the book's own artifacts rather than outside reading; 2.1, 2.3, and 2.6 pick up [AIE] pairings). Part VII does carry read-along pointers, because the loop chapters lean on [RLHF], [S&B], and [MADL] for the algorithm and kernel math even as the labs are hands-on; its per-chapter rows are below.

Part I — How LLMs Actually Work

Chapter	Read-along
1.1 Tensors, autograd, and number formats	[MADL] ch. 2–4
1.2 Tokenization and embeddings	[BLLM] ch. 2
1.3 Attention from first principles	[BLLM] ch. 3; [MADL] ch. 7–8
1.4 The transformer block	[BLLM] ch. 4; [BRM] App. C
1.5 The language-modeling objective	[BLLM] ch. 5
1.6 Where memory goes: training vs inference	[MADL] ch. 8–9
1.7 Sampling and decoding	[BRM] ch. 2; [AIE] ch. 2 (sampling)
1.8 Anatomy of the open-model zoo	[BRM] ch. 1, App. C

Part II — Inference Engineering

Chapter	Read-along
2.1 Prefill, decode, and the roofline	[AIE] ch. 9
2.3 Quantization: theory and formats	[AIE] ch. 9
2.6 Benchmarking without lying to yourself	[AIE] ch. 9

Part III — Evaluation Engineering

Chapter	Read-along
3.1 What an eval is	[BRM] ch. 3; [RLHF] Part 3; [AIE] ch. 3–4
3.2 Metrics and their math	[BRM] ch. 3; [RLHF] Part 3; [AIE] ch. 3
3.6 Judge models	[RLHF]; [AIE] ch. 3 (AI-as-judge)
3.7 The statistics of evals	[CAI]
3.8 Contamination and dataset hygiene	[BRM]; [AIE] ch. 8
3.9 The SDA data pipeline	[DPA] ch. 2–4; [AIE] ch. 8
3.10 From orbital data to verifiable tasks	[BRM] ch. 3; [AIE] ch. 8
3.12 Eval ops	[RLHF]; [AIE] ch. 10

Part IV — Causal Inference for Evaluation

Part-wide spine: [CAI] Parts 1–3.

Chapter	Read-along
4.1 The ladder of causation	[CAI] ch. 1–2
4.2 DAGs and d-separation	[CAI] Part 2
4.3 Confounding, colliders, and selection	[CAI] Parts 2–3
4.4 Identification: backdoor and front-door	[CAI] Part 3
4.5 Interventions on models	[CAI] Part 3

Part V — Reinforcement Learning Foundations

Chapter	Read-along
5.1 The RL problem	[S&B] ch. 3
5.2 Value functions and Bellman equations	[S&B] ch. 3–4
5.3 Bandits, exploration, and sampling	[S&B] ch. 2; [GAIA] ch. 10
5.4 The policy gradient theorem	[S&B] ch. 13
5.5 REINFORCE and variance reduction	[S&B] ch. 13; [BRM] ch. 6 sidebars
5.6 Actor-critic and GAE	[RLHF] ch. 6
5.7 Trust regions: from TRPO to PPO	[RLHF] ch. 6; [BRM] ch. 6
5.8 GRPO	[BRM] ch. 6–7; [RLHF] ch. 6
5.9 RLVR: reinforcement with verifiable rewards	[RLHF] ch. 7; [BRM] ch. 6

Part VI — Post-Training, the LLM Way

Chapter	Read-along
6.1 The post-training landscape	[RLHF] ch. 1–3; [AIE] ch. 2 (post-training); [CUST] ch. 1–2 (the adaptation continuum)
6.2 SFT and instruction tuning	[RLHF] ch. 4; [BLLM] ch. 7; [AIE] ch. 7
6.3 LoRA and QLoRA, mathematically	[RLHF] ch. 4; Part II ch. 3; [AIE] ch. 7
6.4 Reward models and preference data	[RLHF] ch. 5
6.5 Direct alignment: DPO and family	[RLHF] ch. 8
6.6 Reasoning models and inference-time scaling	[BRM] ch. 4–5; [RLHF] ch. 7
6.7 Distillation	[BRM] ch. 8

Part VII — The Loop

Chapter	Read-along
7.1 Unsloth internals	[MADL] ch. 3–4
7.2 GRPO on 16GB	[BRM] ch. 6 (GRPO from scratch), 7.2–7.3 (metrics); [RLHF]; [S&B] ch. 13
7.3 Scorers as rewards	[BRM] ch. 3 (math verifier), 6.6 (reward); [RLHF]
7.4 Reward hacking and Goodhart	[BRM] 7.6 (format reward); [RLHF]
7.5 Data and curriculum for RLVR	[AIE] ch. 8
7.6 Measuring the reasoning delta	[BRM] ch. 3 + App. F (evaluation); [RLHF]

Part VIII — Grounding: Retrieval and Tools

Chapter	Read-along
8.1 RAG over space text	[RAG] (esp. Pt. 2); [LC] RAG systems; [GADP] Pattern 6; [AIE] ch. 6
8.2 MCP tools for live SDA data	[LC] ch. 9 (MCP); [GADP] ch. 7 (MCP); [AIE] ch. 6
8.3 The augmentation arms: what caused the gain	[CAI] Part 3; [RLHF] ch. 7; [GADP] ch. 5–6 (path eval, LLM-as-judge)

Part IX — Burst and Scale

Chapter	Read-along
9.3 When to burst	[RLHF] ch. 6 (infra notes)

Part XI — Beyond the Thesis: Toward a Product

Chapter	Read-along
11.1 Agentic RL: training the model to use its tools	[RLHF] ch. 7; [LC] and [AGENTS] (agent patterns: ReAct, LangGraph, memory, MCP)

(Part X, Assembly, leans on the book's own artifacts rather than outside reading, so it carries no read-along rows.)

Note

When a chapter's read-along pointer and this table disagree, the chapter is the source of truth: update this appendix, not the chapter. This index accretes as chapters are drafted, so pairings for chapters still in stub form may be refined when their read-along blocks are written.

The machine-readable twin

Everything above also lives in book/reading-map.toml, keyed by the same short codes and carrying two things this prose cannot: the Zotero item key for each book, and the chapter pairings as data rather than table cells. Tooling reads the TOML; this appendix stays the place the reasoning lives.

Ten of the fifteen books have a Zotero record and carry the pan-build tag there. That tag is an index, not a priority — it marks "this book is cited by a chapter," nothing more. Reading priority lives on a separate axis entirely, so there is one source of truth for what is due.

Five books are cited here but have no Zotero record at all: [AIE], [RAG], [LC], [GADP], [CUST]. Nothing tracks whether they have been read.

Evals as Rewards