Learning coverage map — every subsystem has a home

This map is the guarantee that the learning corpus covers everything in the project. The ground-up curriculum.md is the spine (10 modules, first principles to self-improving factory). This page is the net: every shipped capability in PROJECT_STATUS.md maps to a learning anchor here, so no subsystem is a black box with no explainer.

How to read it: the Anchor column is where to go to learn that thing. Where a concept has a definitive external source, the anchor stays thin and links out (per the repo’s “lean on external sources” rule) rather than re-explaining.

The spine teaches you to read the whole system; this map is the index that proves nothing was left unread.

1. Foundations — the ground-up spine

The 10-module arc. This is the “learn it in order” path; everything below builds on it.

Reference layer under the spine: llm-mechanics-fundamentals.md (RoPE, GQA, MoE, attention variants) and essential-vs-optimization.md (which math defines the model vs which only makes it fast).

2. Post-training methods and internals

Everything under PROJECT_STATUS.md → “Training and post-training.”

3. Data pipeline

PROJECT_STATUS.md → “Data.”

4. Evaluation

PROJECT_STATUS.md → “Evals.” Module 10 is the concept spine; these are the surfaces.

5. Runtime, serving, and agents

PROJECT_STATUS.md → “Runtime/serving.”

6. Quantization, packaging, and inference optimization

PROJECT_STATUS.md → “Packaging/runtime.”

7. Attention and kernel internals

The math oracle is python_ref/model.py; Session 10 is the concept.

8. Browser / WASM / WebGPU track (completed, parked)

PROJECT_STATUS.md → “Completed/parked learning tracks.”

9. Native Mac / MLX / ANE / Apple Foundation Models

10. Interpretability (completed, parked)

11. Multimodal (VLM) — research/parked

12. The factory loop and methodology

The thing all of the above serves.

13. Strategy and external knowledge

Coverage Guarantee

Every shipped capability in PROJECT_STATUS.md → “Features (shipped)” and “Completed/parked learning tracks” appears in exactly one section above with a learning anchor. The factory-loop framing is the checklist:

target -> data -> post-training -> eval -> package -> report

• target / methodology → §12
• data → §3
• post-training → §2, §7 (internals), §9 (Mac runtime), §11 (VLM)
• eval → §4, Module 10
• package → §6
• report → §4, §12
• serve/run the result → §5
• understand the result → §10 (interpretability), §1 (foundations)

Maintenance rule

This map is guarded by ../../scripts/check_learning_roadmap.py (run via bash evals/learning-roadmap-smoke.sh). When a new subsystem ships in PROJECT_STATUS.md, add a row here pointing to its best explainer before calling the feature done — a capability with no learning anchor is an undocumented black box. When an existing doc is the definitive explainer, point to it (DRY); only write a new learn doc when nothing adequate exists.