diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
new file mode 100644
index 0000000..06ce20f
--- /dev/null
+++ b/CONTRIBUTING.md
@@ -0,0 +1,109 @@
+# Contributing
+
+Thanks for interest in improving this repository. This project prioritizes
+**reproducible experiment infrastructure** and honest reporting over positive
+efficacy narratives.
+
+This guide is for people sending **pull requests**. How you get a branch onto
+GitHub depends on your access (see below). The **integrity / PR expectations**
+apply to external and write-access contributors alike; maintainers may use a
+shorter internal path for trivial fixes.
+
+## Who you are (access paths)
+
+| Role | Typical access | How you open a PR |
+|---|---|---|
+| **External contributor** | No push access to this repo | **Fork** the repo under your account, push a branch to *your* fork, open a PR into `CalculatedContent/nanogpt-experiments` `main` |
+| **Collaborator / write access** | Can push branches to this repo | Create a branch on **this** repo from current `main`, push it here, open a PR into `main` (no personal fork required) |
+| **Maintainer** | Admin / merge rights | Same as write access for reviewable changes; direct commits to `main` only for the project’s usual maintainer practice (prefer PRs for non-trivial work) |
+
+**Fork is not a social rank.** It is only the usual GitHub path when you cannot
+push to this repository. If you already have write access, do not fork just to
+satisfy a ritual.
+
+### External contributor (no write access)
+
+1. Fork `CalculatedContent/nanogpt-experiments`.
+2. Branch from current upstream **`main`** (not from stale feature branches).
+3. Push the branch to **your fork**.
+4. Open a PR against `CalculatedContent/nanogpt-experiments` `main`.
+
+### Collaborator with write access
+
+1. Clone this repository (or add it as `origin`).
+2. Branch from current **`main`**.
+3. Push the branch to **this** repo.
+4. Open a PR against `main`.
+
+Prefer **small, single-theme** PRs (docs, logging, or one behavior change) in
+either path.
+
+## Before you open a PR
+
+From a clean checkout of your branch:
+
+```bash
+# infrastructure smoke only — not a scientific result
+./scripts/run_smoke_test.sh /tmp
+```
+
+Fix any smoke failures you introduced. Smoke success does **not** authorize
+scaling-law, acceleration, or generalization claims.
+
+If your change touches the WW-PGD adapter or projection CSV schema, also run the
+repository CI suite (or the subset your PR description names) and note results
+in the PR body.
+
+## Naming: Level 0 is overloaded
+
+Schema/runner Levels 0–2 (`wwgpt --level N`, `configs/levelN_*.yaml`) are **not**
+the same as the `level_0_baseline/` / `level_0_wwpgd/` packages. Schema Level 0 is
+typically **1 layer · 1 head · width 64**; the baseline packages use **4 layers ·
+4 heads · width 128**. Prefer the package READMEs when working under those folders.
+
+## PR expectations
+
+These expectations apply to **any** PR author (external or write-access), unless
+a maintainer explicitly says otherwise for a given change.
+
+| Do | Don't |
+|---|---|
+| State whether the PR is docs-only, logging, or behavior | Claim WW-PGD efficacy or statistical significance from this stack alone |
+| Keep defaults unchanged unless the PR is explicitly about a default | Fabricate α / invent missing spectral fits |
+| Name any new log field and its definition | Treat smoke or notebook demos as publication results |
+| Link related docs (`docs/SCIENTIFIC_INTEGRITY_POLICY.md`, `docs/CONTROL_ARMS.md`) when relevant | Bundle unrelated refactors with a one-line fix |
+
+Suggested PR body checklist:
+
+```text
+## Type
+- [ ] docs only
+- [ ] logging / schema field
+- [ ] behavior (describe default impact)
+
+## Summary
+...
+
+## Test plan
+- [ ] smoke / CI
+- [ ] no scientific efficacy claim
+
+## Out of scope
+...
+```
+
+## Where to look
+
+| Topic | Path |
+|---|---|
+| Root overview + level naming | [`README.md`](README.md) |
+| Scientific integrity | [`docs/SCIENTIFIC_INTEGRITY_POLICY.md`](docs/SCIENTIFIC_INTEGRITY_POLICY.md) |
+| Control arms | [`docs/CONTROL_ARMS.md`](docs/CONTROL_ARMS.md) |
+| Troubleshooting | [`docs/TROUBLESHOOTING.md`](docs/TROUBLESHOOTING.md) |
+| Baseline package (4L/128) | [`level_0_baseline/README.md`](level_0_baseline/README.md) |
+| WW-PGD package sibling | [`level_0_wwpgd/README.md`](level_0_wwpgd/README.md) |
+
+## License
+
+By contributing, you agree that your contributions are licensed under the same
+terms as this repository (`LICENSE`).
diff --git a/README.md b/README.md
index 4918e4c..32d45d1 100644
--- a/README.md
+++ b/README.md
@@ -2,6 +2,8 @@
 
 This repository runs append-only, paired nanoGPT experiments comparing base optimizers with and without the repository's WW-PGD extension. It provides experiment infrastructure and descriptive analysis; it does **not** ship a scaling-law fit, acceleration conclusion, statistical-significance test, or alpha-generalization result.
 
+See [CONTRIBUTING.md](CONTRIBUTING.md) for fork/PR workflow, smoke expectations, and no-claim norms.
+
 ## Quick start on a local MacBook
 
 Run the repository from its checked-out `main` branch and keep the prepared corpus and results under `/tmp`: