Why training data is durable
Models churn; the data that taught them does not. A well-governed training corpus survives several model generations and underwrites successive releases.
Recognition criteria
Control, identifiability, future benefit, measurement reliability, and a defensible rights basis. Without all five, the corpus is excluded from the registry.
Evaluation sets count too
Curated evaluation sets are often the highest-leverage asset in the AI stack and the least tracked. They sit inside the registry like any other recognized asset.
Questions buyers actually ask.
Only with a defensible rights basis. Most scraped corpora fail at the licensing chain or jurisdictional gate.
As a separate class with its own provenance - the generator, the seed corpus, and the validation harness all become part of the evidence chain.
Weights sit at the model layer, not the data layer. KRYOS scopes the data; the model is a separate analysis.
The vendor relationship becomes the rights basis. The recognition decision turns on what the entity actually controls.