source docs/learn/coverage-map.md · view on GitHub ↗

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.

ModuleTopicAnchor
1Functions, data, parameterssession-01-neural-net-basics.md
2Loss and gradient descentsession-02-gradient-descent.md
3Vectors, matrices, tensorssession-09-tensors.md
4Non-linear nets + backpropsession-03-non-linearities.md
5ML paradigms and scalingsession-04-ml-paradigms.md, session-05-scaling.md
6Tokenization, embeddings, LMsession-06-tokenization-embeddings.md
7Attention and transformer blockssession-10-attention.md
8Training mechanicssession-08-training-mechanics.md
9Post-training: SFT, LoRA, preference tuningsession-07-behavior-learning.md
10Evals, rewards, self-improvementsession-11-evals-rewards.md

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.”

CapabilityAnchor
SFT../training/sft.md, Module 9
Pretraining../training/pretrain.md, Session 8
LoRA / DoRA / QLoRA../lora_guide.md, ../peft_variants.md, ../factory/lora-geometry.md, Session 9
DPO / SimPO / preference tuning../training/dpo.md, Module 9 + 10
Distillation../distillation.md, diversity-driven-small-model-reasoning.md
Evolution strategies (ES)../evolution_strategies.md, castform-rl-finetune.md
RLVR / ReST / GRPOadvanced-llm-training.md, ../GRPO_CLARIFY.md
Optimizers, schedules, stability../optimizers.md, ../galore_and_stability.md, Session 8
NEFTune, z-loss, WSD, LLRD, seq packing, grad checkpointing../training_guide.md, ../gradient_checkpointing_results.md, Session 8
Precision (bf16 / fp8 / mixed)../precision.md, advanced-llm-training.md
Method-vs-recipe discipline../techniques/method-vs-recipe.md

3. Data pipeline

PROJECT_STATUS.md → “Data.”

CapabilityAnchor
Dataset registry / inventory../dataset-inventory.md, ../data_inventory.md
HF integration../hf_datasets_integration.md
GitHub fetcher../github_data_integration.md
Magpie / synthesisadvanced-llm-training.md (data curation), small-model-tool-calling-playbook.md
Tokenizer training, extractor datasession-06-tokenization-embeddings.md, ../tool_call_extractor.md
Traces → data, corrections → data../recipes/from-traces.md
Quality filter, dedupe, reasoning-classify../factory/post-training-factory.md, castform-rl-finetune.md

4. Evaluation

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

CapabilityAnchor
Eval protocol, frozen baselines, gates../factory/eval-protocol.md, Module 10
Eval methodology / broken-eval lessonseval-methodology-2026-06-08.md, eval-matrix-2026-06-08.md
BFCL / tool-calling evaltool-calling-frontier-parity.md, small-model-tool-calling-playbook.md
SQL execution / exact / slices / candidate-choice../techniques/sql-technique-backlog.md, Module 10
lm-eval / HumanEval / MTEB / MILU / tau-bench../lm_eval_integration.md, advanced-ml-systems-eval.md
Router / routed-specialist eval, eval-gate, planner eval../recipes/eval-gate.md, ../recipes/eval_planner.md, ../planner-lock-2026-06-19.md
LLM-as-judge, perplexity, contaminationadvanced-ml-systems-eval.md, Module 10
Leaderboard / reporting../leaderboard.md, ../factory/reports.md

5. Runtime, serving, and agents

PROJECT_STATUS.md → “Runtime/serving.”

CapabilityAnchor
Inference/serving architecture, batching, rooflineadvanced-llm-inference.md
KV cache + paging../kv_cache_optimization.md, advanced-llm-inference.md
OpenAI/Ollama-compatible serve, Continue provider../agent_runtime.md, ../continue_provider.md
Agent loop, tool dispatchmodel-vs-agent.md, ../async_tool_dispatch.md, agent-context-hierarchy.md
Constrained JSON / FSM generation../constrained_generation.md
Speculative decoding / MTP../speculative_heads.md, ../mtp.md, advanced-llm-inference.md
Cost routing / escalation / cascade (AutoMix, ScaleDown)../recipes/automix.md, ../recipes/b25-scaledown.md, agent-context-hierarchy.md
Streaming / long context (StreamingLLM, KIVI)../streaming_llm_kivi.md

6. Quantization, packaging, and inference optimization

PROJECT_STATUS.md → “Packaging/runtime.”

CapabilityAnchor
Quantization (GGUF / AWQ / GPTQ / HQQ), quant theory../quantization_expansion.md, advanced-llm-inference.md
Export to MLX / safetensors / CoreML../recipes/mlx-export.md, ../factory/packaging.md
merge / bake-lora../factory/lora-geometry.md, ../lora_guide.md
Pruning../pruning.md
Specialist packaging / model cards../factory/packaging.md, ../factory/public-artifacts.md

7. Attention and kernel internals

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

CapabilityAnchor
Attention math, transformer blocksession-10-attention.md
FlashAttention-2 forward/backward../fa2_forward_notes.md, ../fa2_backward_notes.md
Online softmax../online_softmax_in_attention.md
MoE / expert routing../moe.md, llm-mechanics-fundamentals.md

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

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

CapabilityAnchor
WebGPU execution model (read before the .wgsl files)webgpu-execution-model.md
Browser GPT training, WASM SIMD, OPFS../browser_notes.md
BPE-in-browser scoring../bpe_browser_scoring.md
Numerics gates / precision drift../precision.md, ../determinism.md

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

CapabilityAnchor
Mac-local mastery map (living agenda)mac-mastery-map.md
Apple on-device Foundation Models — where they fitapple-on-device-foundation-models.md, app-intents-comparison.md
ANE / CoreML research (negative results)ane-research/
Native runtime architecture../../native-mac/ARCHITECTURE.md

10. Interpretability (completed, parked)

CapabilityAnchor
Attention heatmap + logit lens (browser)../interpretability.md, Session 10
SAE, ROME, MEMIT, tuned/logit lens, activation patching../interpretability.md, advanced-ml-systems-eval.md

11. Multimodal (VLM) — research/parked

CapabilityAnchor
Qwen3-VL mRoPE + DeepStack, vision-language attentionqwen3-vl-mrope-deepstack.md

12. The factory loop and methodology

The thing all of the above serves.

CapabilityAnchor
Factory overview + run schema../factory/overview.md, ../factory/run-schema.md
Batch post-training / rollout plan../factory/batch-posttraining.md
Case-study reports, public artifacts../factory/case-study-template.md, ../factory/public-artifacts.md
Attempt history (worked/failed/regressed)../attempt-ledger.md
Owner learning sequence tied to factory work../learning-pipeline.md, ../learning-progress.md

13. Strategy and external knowledge

CapabilityAnchor
Competitive landscape, Mac-first whitespacecompetitive-landscape.md
Reviewed external products / stolen techniques../external-products-reviewed.md, castform-rl-finetune.md
Speech & systems interview-grade topicsspeech-and-systems-topics.md
Papers / reading listexternal-references.md, ../CITATIONS.md
Running questions and tangentsjournal.md

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

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.