Does Assevra execute my agent?
run and scan evaluate captured records. capture executes the command you explicitly provide. The Python recorder can wrap your own in-process execution. None of these automatically replays side-effecting tools from a trace.
Can I run locally?
Yes. Install the package, then use deterministic checks or configure a local judge endpoint. Installation downloads packages. Cloud judges receive the rubric’s input/context/output fields.
Can a scan pass my release?
No. A scan is TRIAGE, and mock judging is SELF_TEST. Declare release dimensions and supply labeled business assertions to obtain release evidence.
What happens when a judge is missing?
Its dimension is skipped and release evaluation is INCOMPLETE. Invalid verdicts or incomplete/tied panels also produce missing evidence, not passing measurements.
Does a suggested action fix the agent automatically?
No. Suggestions guide review. Execute the changed agent again and rerun the same complete suite to verify a fix.
Are integrations native vendor connections?
The SDK and recorder are direct Python interfaces. Other integrations include serialized trace adapters, tool-contract imports, and capture/export recipes. See the maturity table. They do not imply vendor partnerships or validation of your hosted deployment.
Does signing certify reliability or compliance?
No. A signature verifies canonical JSON integrity; a trusted pinned key establishes authorship. Scores describe the supplied cases. Governance mappings are indicative evidence organization, not certification.
Is the core dependency-free?
No. Assevra 0.6 uses jsonschema for complete imported tool contracts. Provider SDKs, Presidio, and signing remain optional extras.
How do I upgrade from 0.5?
Use schema v2 and branch on decision. Required missing evidence blocks release, detector controls are separate, and mock runs cannot qualify as agent release evidence. Read the migration guide.