How it works
How Auvryn runs a model: the ExecutionJob
An ExecutionJob is one run of a validated model on the provider you chose. It is a durable record with a single lifecycle: Auvryn checks compatibility, submits the run exactly once, follows it until it is final, copies the result into its own storage with a SHA-256, and records where it ran and how long it took.
Auvryn is provider-neutral execution infrastructure for AI models. It validates ONNX models, runs them as durable ExecutionJobs on infrastructure it does not own, and returns verified results with provenance and telemetry.
Before anything runs
A model enters Auvryn as an ONNX file. Its SHA-256 is verified on upload and the official ONNX checker inspects it; a model that fails is never offered to a provider.
An optional sandbox test runs the model once in an isolated container, with no network, a read-only root and an unprivileged user, and records the shapes of its inputs and outputs.
The lifecycle
Every ExecutionJob moves through the same states, whatever the provider. Auvryn never moves a job backwards, even when a provider reports a stale state.
- created
- The job exists, tied to one model version, an optional area of interest and the provider you chose.
- validating
- Auvryn asks the provider whether it can run this model: format, operator set, input types, artifact size.
- ready
- Compatible. The job can be submitted.
- scheduled
- Submitted exactly once, with an idempotency key kept for every retry: a retried submission never starts a second run.
- running
- The provider reports the run in progress. Auvryn polls with bounded retries and backoff.
- processing
- The provider finished. Auvryn copies every result artifact into its own storage and verifies its SHA-256.
- completed · failed · cancelled
- Final. Telemetry and provenance are recorded once, at this moment, and never rewritten.
Succeeded is not the same as correct
A provider saying a run succeeded only means it finished. For reference workloads with a known expected output, Auvryn compares the returned values with the expected ones itself, within a stated tolerance, and records that verdict separately, as a figure it derived.
When something fails
Failures are normalized: an incompatible model, a provider that rejects the submission, a run that fails on the provider, a result that cannot be retrieved or does not match its checksum. Transient provider errors are retried a bounded number of times; a job never hangs forever and a succeeded run is never executed twice.
Where jobs run today
You choose the provider for every job. The Sample Environment is simulated and labelled as such. Auvryn has also run ExecutionJobs on real external edge hardware, under internal evaluation and not yet offered to customers. Orbital providers are not connected yet.
Related
Facts reviewed 2026-10-07.