Cite a dataset by its hash, not a URL
Publish a dataset and it becomes a content-addressed, tamper-evident, versioned, citeable artifact. The citation is the SHA-256 of the content — so a paper cites at1://<hash>instead of a link that rots, and anyone can resolve it byte-identical or the resolve refuses. Reproducibility isn’t a promise; it’s the address.
Publish once, cite forever, resolve anywhere
at1 dataset publish nyc-taxi-2024.csv --name nyc-taxi-2024 # nyc-taxi-2024 v1 — 1.9 GB # cite: at1://b4eb94e47a78710… at1 dataset resolve b4eb94e47a78710… -o data.csv # OK: resolved b4eb94e47a78710… -> data.csv (integrity verified) at1 dataset log nyc-taxi-2024 # v1 b4eb94e47a78710… 1.9 GB (first) # v2 a91c3f00d2e4a19… 1.9 GB (prev b4eb94e47a78…)
The hash is the address
A modified dataset cannot resolve to the same hash — resolve re-checks the bytes and refuses on mismatch. Integrity is structural, not a claim.
Versioned & citeable
Re-publishing a name appends an immutable version linked to its predecessor — an append-only lineage. Cite an exact snapshot in a paper or a pipeline; it never drifts.
Queryable-in-place
Datasets are stored compressed and verified; the same container is queryable in place, so a cited hash is also a dataset you can run SQL against without downloading it whole.
For reproducible ML and research data, data marketplaces, and pipelines that need a dataset to be exactly what it says it is. Pay-as-you-go: first 100 publishes/month free; resolve and verify are always free.