Package a task and export results

Choose what you are handing over

A training checkpoint, deployment package, prediction journal and result archive serve different purposes. Preserve the object required for the consumer’s next operation.

Package the selected execution closure

Select the reconstruction entrypoint, application code and resources needed by the task. A package does not copy the entire training project by default. A frozen host reference remains separate unless explicitly selected. Consumers need compatible dependencies and the declared resources.

python workflow.py package
python workflow.py runtime
python workflow.py runtime --resume

Define the result layout

ResultExportConfig selects your installed writer, revision, options and optional validator. The writer declares named relative files per original ID. Export verifies the original ID population exactly and preserves native JSON/file outputs as provenance.

The destination must be fresh. Missing results, duplicate IDs, file errors or a failed validator prevent partial publication and leave the source journal intact. A new layout revision creates a new export from the same committed results.

Verify the archive after moving it

Inspection validates the complete inventory, bytes, digests and declared result population without importing model or writer code. Instructions and unresolved checks remain included. Export neither runs inference nor sends results to another system.

python workflow.py export
python -m arti recipe inspect-results artifacts/results.zip