Document library — where to look
A labeled map of the retained learning, reference, factory, evidence, and historical documentation.
Use this map when you need a document outside the current learning session. The sidebar shows the primary entrances; site search reaches the full retained corpus. Existing document URLs and Markdown mirrors remain available.
Learn and build
| Shelf | Use it for | Start here |
|---|---|---|
| Current course | The owner’s 13-week inference route and its evidence contract | Course workspace, weekly route |
| One active task | Exercise, checkpoint, and actual delayed recall dates | Next Session |
| Foundations | Prerequisites from functions and tensors to attention and evals | Ten-module curriculum, ordered lessons |
| Practical labs | Runnable paths, artifacts, exercises, and mastery gates | Artifact journey, learning pipeline |
| Model and systems concepts | Architecture, inference, training, and Mac-to-cluster boundary | Learning index, Mac mastery map |
Apply and evaluate
| Shelf | Use it for | Start here |
|---|---|---|
| Factory contract | Target → data → post-training → eval → package → report | Factory docs |
| Methods and recipes | Distinguish a technique from a runnable, gated recipe | Technique registry, recipe index |
| CLI and integrations | Commands, local workflow, and external interfaces | CLI reference, integration example |
| Data and models | Dataset provenance and retained model artifacts | Data inventory, model card |
Inspect the evidence
| Shelf | Use it for | Start here |
|---|---|---|
| Outcomes | Final dispositions and failure explanations | Attempt ledger, public artifacts |
| Performance | Measurements, harness design, and method limits | Benchmark harness design, Mac decode baseline |
| Research | Benchmark and landscape findings, with sources | Inference benchmark survey, external products reviewed |
| Audits | Coverage and validation boundaries | History coverage, docs quality |
Read with historical context
| Shelf | Use it for | Start here |
|---|---|---|
| Status and admission | Closed-lab state and the rule for a fresh experiment | Project status, closure and next rule |
| Deferred specifications | Acceptance criteria for named historical or parked ideas | PRD index, priority map |
| Parked lanes | Work intentionally paused | Parked lanes |
| Archive and session records | Historical plans, transcripts, and dated handoffs | Archive status, document status registry |
| Original walkthrough | Python → WASM → WebGPU guided tour | Legacy walkthrough |
Historical TODOs and PRDs are reference material. New model work still needs an owner-selected question and the fresh-experiment gate.
Retained evidence directory
These direct links keep older evidence discoverable without turning the sidebar into a full archive. They are historical records, not an active backlog or new claims about model quality.
Performance receipts
- First Worked Run —
posttrainllm bench - Cold-start bundle — results
- CPU speedup bundle — measured results
- CPU utilization research — what we’re leaving on the table
- Data-side performance levers
- YOCO — “You Only Cache Once” worked example
Research and audit receipts
- Learnable cross-stream attention — adopt or skip?
- Indic-language evals — MILU + IndicGenBench
- Wave 2.5 kernel audit — BUILD / DEFER / DROP
- Mac SLM agentic leaderboard v0
- Feature audit — 2026-05-31
- Test Coverage Goals — posttrainllm
- GitHub Corpus Parity Decision
Factory and specialist references
- Foundry Evidence Contract — posttrainllm
- Hugging Face Artifact Storage
- Qwen3-4B ReST Fused
- Recipe — Pace planner specialist
- Recipe: distill a big model into a small local specialist
- A1 — first tool-calling specialist
- B1 — second specialist: text-to-SQL
- parakeet.wgsl browser smoke
- Needle 2 baseline review
Dated sessions and retrospectives
- Session — eval-first prep before the 2-day training window
- Strategy session — what posttrainllm actually is, what’s possible on a Mac
- Specialist Embedder PRD Blocker
- vllm-mlx Investigation
- Session — 2026-06-08, the factory-validation night
- Strategy session — the market landscape and the Mac-first wedge
- Session — from frontier-parity distillation to the self-improving loop
- Step-back: what we have, the wall, and the ROI menu
- Vercel Eve and the local-agent wedge
- Drilldown — the experiments left untried, run to bedrock
- PostTrainLLM — retrospective and forward plan
- Pace handoff — posttrainllm deliverables status (2026-06-10)
- Planner lock - 2026-06-19
- PostTrainLLM — Working Session Q&A Log
- Session retrospective — 2026-05-31
- Session Retrospective: posttrainllm, the Long Working Session
- First specialist run —
toolcall-v1findings - v11 baselines — committed 2026-06-09
- WWDC 2026 (June 8–9) — impact on posttrainllm + Pace