Developer docsCompute · Arena · API

Trust, failures, and operations

Understand evidence labels, permissions, retries, and what to inspect when a job fails.

Operators and platform engineers · 8 min · Updated
Markdown

Read the evidence label first

Compute result trust labels
LabelMeaningMay claim
provider_reportedA provider returned a result that has not passed the full verified projectionProvider-reported facts with disclosure
provider_verifiedAttempt/spec/package/manifest/attestation/signature/evaluator checks passedOnly the exact verified private result; not a hardware claim
Simulation result recorded by HIMA reviewed simulator and evaluator ran under a recorded execution boundaryOnly that recorded simulation environment and profile

Failure ownership matters

Use the failure class to choose the next action.
ClassExamplesNext action
PolicyUnapproved profile, digest drift, unsafe path, missing grantFix configuration/approval; do not retry unchanged
PackageEntrypoint failure, missing declared output, contract mismatchFix or republish a new immutable package
SimulatorAsset load, physics/task crash, invalid trajectoryInspect adapter/environment and reproduce locally
EvaluatorMalformed/non-authoritative metric outputFix the evaluator separately; candidate output stays frozen
InfrastructureProvider timeout, preemption, upload/callback transportRetry/reconcile according to retryable flag and cost policy
CancelledOwner cancellation inside the allowed windowRetain the lifecycle record; resume training only from a published checkpoint

Operator checklist

  1. Open the job and identify status, stage, attempt, provider job ID, spec digest, package entrypoint, and last immutable event.
  2. Confirm whether the output is provider reported, provider verified, or reproduced by HIM in simulation.
  3. Inspect bounded logs and environment identity without exposing secrets or private source paths.
  4. Check artifact manifest, attestation, evaluator digest, signature key ID/state, callback identity, and owner-only publication projection.
  5. Choose cancel, retry, resume, or package/config fix based on the failure owner and retryable flag.
  6. Record the incident and do not publish a score while its evidence is uncertain.