Artifacts & serialization

ARTI uses SafeTensors for weights and JSON sidecars for metadata and integrity checks.

Save and load

saved = arti.save(layer, "layer.arti.st")
loaded = arti.load("layer.arti.st", model=fresh_layer)

print(saved.weights_sha256)
print(loaded.missing_keys, loaded.unexpected_keys)

Format compatibility

ARTI 1.x reads compatible format-version 1 artifacts produced by the pre-public 0.x line. Normal .arti.st loading does not depend on Python pickle.

Integrity is not identity

The JSON lock sidecar records hashes that detect modification relative to that lock file. A matching hash is not a publisher signature.

Obtain models and weights from trusted sources, and pin external model and dependency revisions according to your deployment policy.

Migrate legacy tensor-only state

arti.migrate_pt("legacy-state.pt", "layer.arti.st")

Migration uses PyTorch’s restricted weights_only=True loader. Full pickled Python objects are not supported.