posttrainllm Docs
This is the canonical entrypoint for posttrainllm documentation.
posttrainllm is a Mac-local specialist factory. The docs should let a new reader answer five questions without asking the project owner:
- What is this project?
- What have we tried?
- What worked, failed, or is still blocked?
- What external products/papers/blogs changed the roadmap?
- What should we build and learn next?
Golden Path
Read in this order:
| Step | Doc | Purpose |
|---|---|---|
| 1 | ../PROJECT_STATUS.md | Current state, active product thesis, shipped/parked surfaces, active gaps |
| 2 | NEXT.md | Active queue only; what to do next and what not to touch |
| 3 | factory/README.md | Factory contract: run schema, reports, evals, packaging, public artifacts |
| 4 | techniques/README.md | Method-vs-recipe registry; how we choose model-improvement tactics |
| 5 | techniques/audit-inventory.md | Row-level treatment of the broad 2026 technique audit |
| 6 | attempt-ledger.md | What worked, failed, regressed, or remains untried |
| 7 | external-products-reviewed.md | Competitors/startups/papers/blogs reviewed and what we stole or rejected |
| 8 | learn/curriculum.md | Ground-up learning roadmap from first principles to self-improving specialists; learn/coverage-map.md indexes every subsystem to a learning anchor |
| 9 | learning-progress.md | Measured progress through the owner learning pipeline |
| 10 | learning-pipeline.md | How ground-up learning attaches to the factory lab |
| 11 | factory/public-artifacts.md | Public artifact inventory, blockers, and release posture |
| 12 | history-coverage-audit.md | Exactness boundary for normalized vs unnormalized historical attempts |
| 13 | exactness-completion-audit.md | Completion proof for the docs exactness pass |
| 14 | docs-quality-audit.md | Honest audit of what is world-class now and what still is not |
| 15 | doc-status.md | Active/reference/parked/superseded labels for major docs |
Current Project Loop
target -> data -> post-training -> eval -> package -> reportEvery active task should improve one of those stages. If it does not, it belongs
in docs/parked/, docs/prds/, or the learning corpus.
Documentation Standard
For every meaningful model/factory attempt, the docs must record:
- target and frozen eval
- recipe, not just method
- data source and provenance
- baseline result
- candidate result
- slice metrics
- trace review or failure taxonomy
- failure reason confidence and evidence sources
- performance where feasible
- decision:
ship,retry, orreject - next action
For every external product or paper that changes the plan, the docs must record:
- source
- useful technique
- local translation
- status: adopted, scaffolded, tried, rejected, or parked
- next smallest experiment
Active vs Reference vs Archive
| Category | Meaning | Where |
|---|---|---|
| Active | Current work queue and gates | PROJECT_STATUS.md, docs/NEXT.md, docs/factory/, docs/techniques/ |
| Evidence | Attempt results, reports, artifact status | docs/attempt-ledger.md, docs/specialists/, docs/factory/public-artifacts.md, runs/*/report.md |
| Learning | Owner curriculum and concept explanations | docs/learn/, docs/learning-pipeline.md, docs/training/ |
| Reference | Broad mechanics, recipes, historical plans | docs/recipes/, docs/roadmap/, docs/prds/, docs/PLAN.md |
| Archive | Superseded or moved material | docs/archive/, docs/parked/ |
If two docs conflict, prefer current source in this order:
PROJECT_STATUS.md
-> docs/NEXT.md
-> docs/techniques/
-> docs/factory/
-> run report / artifact entry
-> older PRD/PLAN/session notesSee doc-status.md for the status label of each major docs
surface.