Support scope

Package and component status

ARTI 3.3.0 is the published package snapshot used by these pages. Neural components retain their own lifecycle and contract addresses. A stable package release does not silently change an Alpha component's lifecycle or erase the exact schema stored in an asset.

Documented environments

The distribution requires Python 3.10+ and PyTorch 2.2+. The release matrix documents Python 3.10, 3.11 and 3.12. Optional model, training, JAX and Flax integrations have additional dependency and version requirements. A successful local example does not broaden that matrix.

Keep execution and quality separate

A completed run, a finite gradient and an exported file establish different software observations. Task acceptance needs the application's actual targets, held-out inputs and evaluation protocol. These pages make no general model-quality or GPU-throughput promise.

• Preserve raw native predictions and their request identities.
• Do not replace the task outcome with an internal diagnostic loss.
• Treat imported external evidence as attributed evidence, not as a local reproduction.

Application episode state

EpisodeRunConfig supports declared policy/environment callbacks, finite JSON application state and structured traces. The native model profile is read-only per policy step. Arbitrary native mutable model persistence, binary state codecs and physical-effect rollback are outside that profile.

Local runtime enforcement

The included runtime is a trusted local worker. Its direct-worker deadline has a recorded scope; it is not process-tree isolation or a memory, CPU or GPU cap. Offline network hints do not block sockets. Required unsupported enforcement prevents execution, and container/provider qualification needs its own backend evidence.

Recovery requires the declared state

Only complete supported checkpoints and committed result parts provide the described continuation guarantees. Unfinished callbacks can require another invocation. Ambiguous external actions are not automatically repeated, and local recipe recovery is not a claim of arbitrary distributed or sharded recovery.