docs/PUBLISHING.md

Publishing checklist

The repository intentionally separates source publication from package/model publication.

Hugging Face

1. Review git diff --check, tests, build artifacts, and validation evidence.
2. Create dougvk/Unlimited-OCR-RDNA4 as a private model repository.
3. Push the reviewed initial commit.
4. Verify the model card, source tree, package artifacts, and remote hashes.
5. Make the repository public only after those checks pass.
6. Tag v0.1.0 only after the clean-environment gate in docs/VALIDATION.md is reproduced.

PyPI (deferred)

Do not publish v0.1.0 to PyPI. The supported installation contract is the repository's hash-locked bootstrap, while an installed wheel cannot yet reproduce that bootstrap without returning to the source checkout. Building and running twine check remains a packaging gate, not authorization to upload. A later PyPI release should bundle an equivalent bootstrap entry point and independently verify a fresh wheel installation first.

Do not upload or duplicate Baidu's checkpoint. The repository contains the runtime source and packages and links to the original baidu/Unlimited-OCR weights.

The repository should be classified as documentation for a runtime/integration, not as a runnable Transformers pipeline, new model, checkpoint, fine-tune, or kernel port.

GitHub (optional mirror)

If a GitHub mirror is created later, update the project URLs consistently and enable private vulnerability reporting.

Upstream

After publication:

1. open a concise Baidu discussion or documentation pull request with the verified configuration and evidence;
2. offer the gfx1201 bootstrap and validation findings to AIwork4me/Unlimited-OCR-ROCm;
3. avoid claiming universal ROCm or all-RDNA support until independently reproduced.