Every hour, Crovia fetches the public model card of a rotating set of Hugging Face models, runs one published predicate over the bytes — does this card disclose its training data? — and signs what it saw. The hour is opened by a public randomness beacon (drand) and closed by a Bitcoin anchor (OpenTimestamps). This dataset is that log, byte for byte, as served these bytes (hash and length), and the predicate returned this result. result: false means no training-data disclosure was found on that surface in that hour — not that the provider disclosed nothing anywhere, not fraud, not bad faith, not a quality grade. Hours not observed do not count toward anything. See the and the LACUNA page.…