# Public artifact revisions

These receipts connect the same frozen v0.2 experiment across GitHub, Hugging Face and this research projection. They verify byte identity and publication history; they do not independently authenticate the simulator or qualify a physical robot.

- [Final revisions and hashes](final-revisions.json) identify the GitHub prerelease, unchanged model, MIT dataset card revision and derived recorded replay.
- [Initial Hugging Face upload](initial-upload-receipt.json) records the complete model, evidence dataset and original recorded projection, including anonymous readback and preserved Space files.
- [Dataset card overlay](dataset-card-receipt.json) changes only publication metadata. The original scientific README remains byte-identical under `docs/publication/source-README.md`; use its explicit path mapping when checking the original scientific seal at the final dataset revision. The initial dataset commit also retains the original root inventory directly.
- [Native-resource Space overlay](native-space-receipt.json) records the regular-Git bundle preflight, append-only recorded cases, atomic frontend activation and anonymous final readback. Dataset evidence URLs pin the final card revision. Both public hosts use the same derived viewer seal.
- [GitHub asset readback](github-assets.json) records anonymous checks of the installation kit, wheel and checksum file. Use a fresh virtual environment: experimental wheels still share package version `0.1.0`.
- [GitHub README exposure](github-readme-receipt.json) records the exact prepared content read back at the new main-branch commit. The frozen research tag remains separate.

The complete-study seal and original browser-projection seal remain separate. The smaller viewer does not contain the entire dataset. Hosting changes affect transport and evidence links, while the original simulation states, geometry, model, controls and scores remain unchanged. The browser replays saved states; the local CLI runs fresh physics.

The measured result remains **6/24** strict successes for feedback, **8/24** for the blind learner, **5/24** for the frozen primitive, **24/24** for the authored teacher and **0/24** for open jaws. Feedback succeeds in **0/12** scheduled jaw-interruption cases. No live-feedback advantage, post-grasp recovery or physical transfer is established.

Original DVIDIA code and authored evidence are MIT licensed. Upstream robot and rendering components retain their included notices. The [paper](https://research.dvidia.org/papers/learned-piper-grip-v02/) documents the benchmark and its limits.