Skip to content
This repository was archived by the owner on Jul 23, 2026. It is now read-only.

CONTRIBUTING.md

Latest commit

 

History

History
56 lines (37 loc) · 4.32 KB

CONTRIBUTING.md

File metadata and controls

56 lines (37 loc) · 4.32 KB
Raw

Contributing to CodeWeave

Thanks for your interest — whether you want to use CodeWeave on a codebase of your own, improve the engine, or build on the PyTorch proof run in this repository, there's a place for you here.

First, know which repo you want:

  • This repository (CodeWeave-PyTorch) is a proof run: the CodeWeave engine applied end-to-end to PyTorch (CPU), with the baseline, verdicts, and full audit trail committed. Contribute here if your change is about this run — the results, the analysis, the docs, or reproducing it.
  • The engine lives at HighTech-Innovators/CodeWeave. Contribute there if your change is to the pipeline itself — workflows, prompts, the verdict statistics (ab_compare.py), or portability to new targets.

If you're unsure, open a Discussion and we'll help you place it.

Ways to contribute

For new users

  • Try the Quickstart against your own target and tell us what broke — reproduction reports on a new target are especially valuable, because portability is only verified on PyTorch so far.
  • Improve the docs: anything that confused you on first read is a bug in the README.

For contributors

  • Reproduce a verdict. Re-run a Phase 7 cycle and confirm (or challenge) a result in docs/RESULTS.md. The opt-1 (addmm) result carries an explicit drift caveat and needs a confirming re-run — that's a great first contribution.
  • Fix a limitation. See Known limitations in the README.
  • Extend a gate or tool. The _tools/ helpers (ab_compare.py, diff_fuzz.py, op_microbench.py) are the measurement substrate.

For academics & researchers

  • The methodology is documented in executive-summary.md: correctness-gated-before-measured changes, a significance-and-noise-floor verdict, drift-controlled paired A/B, and a Holm-Bonferroni family-wise correction.
  • All raw artifacts are committed under integration-test/reports/ and proof/ for independent re-analysis.
  • If you use this work, please cite it — see CITATION.cff or the "Cite this repository" button in the sidebar.
  • Found a flaw in the statistical design? Open an issue or Discussion — rigorous critique is the most useful thing you can send.

Ground rules that keep the project honest

CodeWeave's core invariant is that no result is trusted on assertion — it must survive the correctness gate and the statistics. Please preserve that in any contribution:

  • Generators propose; independent validators dispose. Don't let a component certify its own output.
  • ab_compare.py is the single source of truth for every statistic. Don't hand-edit a verdict JSON or quote a number that isn't traceable to a committed artifact.
  • Correctness before performance. A change must pass the differential-fuzz gate before any timing claim.
  • Per-iteration metrics, never wall clock (the workload is a fixed-time hot loop).

Development workflow

  1. Discuss first for anything non-trivial — open an issue or Discussion so effort isn't wasted.
  2. Branch from main with a descriptive name (e.g. docs/..., fix/..., results/...).
  3. Make focused commits with clear messages explaining the why, not just the what.
  4. Don't commit generated or private material — src/ is untracked by design, and local/private docs are .gitignored.
  5. Open a pull request describing the change, its motivation, and (for anything touching results) which committed artifact backs each number.

Reporting bugs & requesting features

Open an issue with: what you expected, what happened, the target/config you ran against, and the relevant artifact or log from proof/.

Security

Please do not file security issues in public. See Security for how to report privately. [ADD SECURITY CONTACT / SECURITY.md]

License

By contributing, you agree that your contributions are licensed under the project's GNU General Public License v3.0 or later (see LICENSE).