Assets and preparation

Capture the declared file closure

DatasetSnapshot retains records, explicit asset bindings and content checksums. Logical assets can share identical copied content. The application chooses IDs, roles, splits and group rules; snapshot creation does not infer labels, codecs or tensor ranks.

Reuse preparation by identity

AssetTransformConfig declares the installed transform, its revision and options. Cache keys combine that declaration with source content and metadata. A retry verifies committed output files and prepares missing entries. The transform must be deterministic and asset-local for that identity to remain meaningful.

• Change the declared revision when decoder code or dependencies change.
• Represent seeds and other deterministic options explicitly.
• The supported cache is local with one writer; it does not provide distributed coordination or automatic code fingerprinting.

Save the supported state

ARTI model assets use SafeTensors with validated metadata and locks. Learned weights, physical resources and training checkpoint state have separate ownership. Construct compatible custom modules explicitly when loading; a manifest does not import arbitrary Python classes on your behalf.

import arti

# model and restored_model are compatible application-owned instances.
arti.save(model, "state.arti.st")
arti.load("state.arti.st", model=restored_model, map_location="cpu")

Inspect compatibility and provenance

Component contract addresses, concrete state schemas and package versions answer different compatibility questions. Shared ownership records refer to actual shared parameter objects, not merely equal values. Hashes detect changes relative to a declared lock; they do not certify publisher identity.

Select the consumer's closure

Include only selected inference resources, reconstruction code and runtime declarations in a task package. Frozen base weights remain distinct from adapter assets. Bundled wheels are explicit offline inputs and are not automatically installed. A portable directory still requires a compatible consumer environment.