PACINFRAX · PRODUCT

Plan asynchronous inference with explicit completion windows.

Define queued inputs, output handling and failure recovery before submitting a batch.

Your decision

Decide whether delay-tolerant inference can move outside the interactive path.

Bound input size, job count and output retention; classify data before choosing upload or artifact storage.
Specify partial-result, duplicate-item and failed-item behavior, with a per-job cost ceiling and reconciliation process.
Cancellation or a disconnected client does not prove compute stopped. Define terminal evidence before releasing an obligation.

Batch submission, queues and result downloads are not enabled. Current text API documentation does not imply a batch API.