How dataset source facts are checked

The refresh job requests the per-dataset API endpoint. It reads an explicit cardData.license value first, then an explicitlicense: tag. Some gated cards put the license in their access agreement; in that case the job accepts only a named license or an official license link in that text. Size comes fromcardData.size_categories or its matching tag. Downloads and creator come from the API record. Papers come only from explicitarxiv: tags.

The generated JSON is committed with the application. Page builds do not call Hugging Face, so a network outage cannot change a build or silently replace source facts. A later refresh can update the snapshot and make the resulting diff reviewable.

A license name is reported as source metadata, not as a complete rights opinion. A license marked non-commercial is labeled as a commercial-use restriction. Other licenses still need review of the linked terms, upstream sources, and the intended use. When no license is provided, commercial use is not established.

Size categories and download counts are discovery facts, not quality scores. Before using a dataset, inspect the current card and files, confirm the relevant modalities and formats, and test a representative sample. Related pages are alternatives for source review; shared tags do not prove that two datasets are equivalent.

Robot demonstration quality is not one score. This methodology keeps automated facts—loadability, schema drift, timing, missing/NaN values, stream lengths, distribution, duplication, and leakage—separate from human judgment about task validity, strategy, safety, and rights. Influence functions and mutual-information estimators are cited robotics-primary curation signals; neither replaces structural validation or a human decision.

TrueLabel analysis is limited to the operational gates and the automated-versus-human split. Source-reported methods retain their exact scope and limitations. This is a transparent demonstration curation record for high-quality robot data review, not a field-canonical grade or a promised model result.

Read the versioned mixture recipe or download the mixture and quality ledger as JSON.

A structural pass does not prove task usefulness. A curation score depends on the model, estimator, and validation set. A repository license does not establish dataset consent, redistribution, or derived-model rights. Keep those unknowns visible, retain reviewer reasons, and evaluate alternate retained sets on the same held-out target embodiment.