Artifacts and Model Formats

Artifacts are the contract surface that quant deployments actually run.

A quant job is useful because it produces an artifact. A release matters because it points to a specific artifact. A deployment is meaningful because it runs that artifact under controlled runtime settings.

Artifact targets

PyP currently centers quant around a small set of artifact targets:

python_bundle

Best for:

• custom Python logic
• lightweight deterministic models
• simple numpy-driven parameter bundles

Typical behavior:

• artifact.json contains strategy code, config, metrics, and model payload
• live runtime can execute it directly in the Python worker path

joblib

Best for:

• exact scikit-learn style artifact preservation
• classical tabular model workflows

Typical behavior:

• artifact.json stores metadata
• a real model.joblib is stored separately
• exact runtime support depends on the runtime family implementation

onnx

Best for:

• exact preserved model files
• framework export paths that target a portable runtime format

Typical behavior:

• artifact.json stores metadata and model URI
• a real model.onnx is stored separately
• heavier runtime paths may use container-backed execution

Why artifacts matter

Artifacts solve three different problems at once:

• deployment reproducibility
• release versioning
• simulation integrity

Without an artifact, "the model" is just a vague training result. With an artifact, PyP can treat the output as an actual deployable unit.

Artifact metadata

A healthy quant artifact should carry:

• artifact format
• runtime type
• artifact target
• metrics
• config
• model storage mode
• created time

That metadata is what lets the router, release flow, and deployment pipeline understand what they are dealing with.

Best practice

• choose the simplest target that preserves what you need
• use python_bundle when lightweight custom logic is enough
• use joblib when exact sklearn-style preservation matters
• use onnx when the exact model file must remain portable and intact