Skip to main content
A blocked or inconclusive result is not one universal failure. Identify the layer before changing a threshold, rerunning a job, or asking another model. Exact error envelopes vary by API and layer. Use the returned machine-readable code and the matching contract rather than searching only for a generic “failed” message.

​
Do not treat a retry as new independent support

Retrying can recover execution. It does not turn one input into a new independent example. Stability repetitions should remain attached to the original case or higher-level document/conversation group.

​
Do not change the question after reading the answer

A comparison can become invalid when the reference labels, application output, evidence construction, or policy changes between arms. Pin a new comparison when the protocol changes. Do not rewrite the original result to make its identities match.

​
Check semantic consistency before interpretation

If one field says the final composition passed while child provenance still requests escalation, inspect the later fallback attempt. If an early stability run passed and a later one failed, the early result does not settle the case. A contradiction between the terminal result and the saved comparison is a reporting/integrity issue. Preserve it for investigation rather than selecting whichever outcome allows progress.

​
Check which client and deployment you are using

Source support, a generated REST operation, a convenience SDK method, a CLI workflow, and a released package are different levels of availability. The feature inventory explicitly tracks gaps. Avoid undocumented web requests or invented CLI flags as workarounds.

​
Gather an effective support bundle

Include the source/deployment identity, evaluator release, validation and measurement revision IDs, policy identity, final outcome, relevant child-attempt states, and reason codes. Add only the minimum redacted case evidence required to reproduce the issue. Do not include provider keys, session tokens, or unrestricted customer payloads. Continue with Reading evaluation evidence, Qualification, and Feature status.