# Release audit (2026-09-02)

This audit covers the clean public release that is mirrored in both GitHub repositories.

## Addressed in the clean subset

- Removed cluster-local model, dataset, cache, and output defaults from the public subset.
- Added explicit model/repository arguments for the LLaDA2.0-Uni adapters.
- Fixed bare output filenames in the MMaDA trainer (`Path(...).parent` now handles `delta.pt`).
- Kept model weights, benchmark data, generated artifacts, logs, result tables, and nested Git metadata out of `open_source/`.
- Added a network-free syntax/path/schema checker and an offline HTML project page.
- Added license, third-party notices, and a responsible-disclosure policy.
- Deliberately omitted the LaTeX/source archive from the public release; `paper.pdf` and web-ready figures are retained.

## Remaining decisions for the authors

1. Confirm that the paper figures and qualitative examples in `assets/` may be redistributed and replace the placeholder copyright holder in `LICENSE` if needed.
2. Add the ACL Anthology URL and page numbers to the BibTeX entry once the EMNLP 2026 proceedings record is published.
3. Add official upstream URLs and commit hashes for LLaDA-V, MMaDA/MAGVITv2, and LLaDA2.0-Uni.
4. Decide whether benchmark assets can be released under their own terms. The current subset assumes they remain private or are downloaded directly from their maintainers.
5. Re-run the full benchmark from a clean environment and publish a machine-readable manifest (model revision, data revision, seed, preprocessing, judge, and confidence intervals) rather than copying old logs.
6. Add dataset versions, access instructions, and preprocessing details if exact reproduction is expected.

## Working-tree risks not copied

The surrounding working tree contains nested `.git` directories, extensive experiment logs/results, and documents with author-machine paths and contact metadata. Publishing the repository root recursively can expose that information and may also redistribute third-party datasets or model artifacts. Publish only the curated subset after checking `python open_source/scripts/check_release.py`.