CDIF answers a different question from Croissant. Croissant says, “Here is how this dataset is structured.” CDIF says, “Here is how this dataset participates in a larger interoperable data ecosystem.” That distinction matters because an AI Navigator must operate both inside a local lakehouse and across institutions, domains, catalogs, and communities of practice.
The first principle of CDIF is interoperability. A local system can be precise and still be provincial. If each repository describes discovery, access, rights, variables, and provenance differently, agents cannot move responsibly across domains. CDIF gives Querygraph a way to publish local assets in a language that other FAIR data systems can understand.
This is another answer to Big AI. Instead of asking a model to infer cross- domain meaning from whatever text happens to be nearby, CDIF makes the interoperability layer explicit. Discovery, manifest, data description, access, rights, vocabularies, integration, universals, and provenance become structured profiles rather than hidden prompt assumptions.
In Querygraph, CDIF is the publication projection over the
Croissant-grounded asset. The Rust module cdif.rs projects
datasets into profiles for discovery, manifest, data description, data
access, access rights, controlled vocabulary, integration, universals,
and provenance. Those profiles are not bureaucratic checkboxes. They are
the handles that make cross-domain AI possible.
Imagine a resilience analyst asking whether fiscal fragility and energy insecurity overlap in vulnerable communities. Finance tables and energy survey tables do not naturally speak the same language. CDIF helps Querygraph describe the asset in a way another system can discover, compare, cite, and connect. The CDIF projection gives the navigator publication-grade metadata while Croissant remains the close-up record-set description.
CDIF lives beside the lakehouse, not above it as an abstract
aspiration. For each dataset Querygraph loads, the sidecar
semantic/cdif.json travels with the corresponding prepared
data and Croissant sidecar. The validator checks that these semantic
artifacts remain shaped correctly, because stale metadata is worse than
no metadata: it gives an agent confidence in the wrong map.
Textbook rule: Croissant makes a dataset locally legible; CDIF makes it federation-ready. Querygraph needs both because the navigator must be precise inside one Sail warehouse and intelligible across many catalogs.