Third-party materials

The MIT license covers this project's original code and explanatory documentation. It does not grant rights to third-party benchmark text or model weights.

• Google DiffusionGemma: weights are downloaded from Hugging Face by the user, subject to Google's model access and license terms. No weights are bundled.
• TypeSafe AI public evaluations: typesafe-public-evals/*.json contains the public viewer's selected inputs, questions, reference answers and saved outputs. manifest.json records source URLs, retrieval date and SHA-256 hashes. These are attributed evaluation materials, not a new license grant. The corresponding result receipt also includes saved reference outputs.
• Every TypeSafe Parallel Judgment Lab: the original downloaded ZIP, notebook and app source are not included. Import a copy obtained from Every with scripts/import_every.py. The included result receipts retain the prompts, labels, reference scores and source hashes needed to inspect our two experiments; those underlying benchmark materials retain their original rights. The source bundle describes support examples as synthetic and Slack-derived material as de-identified summaries.

Saved model outputs and references are not independently verified human ground truth. This project is not affiliated with or endorsed by Google, Every or TypeSafe AI.