How to prove what an AI agent did, and what proof cannot cover

Three separate claims

• Capture: a runtime integration observed the execution and recorded it as it happened.
• Verification: the stored artifact still hashes to the value it was sealed with, so it has not been altered.
• Certification: an independent NexArt Node attests to an eligible stored artifact, unchanged.

The order matters

Certification adds nothing to a record that was never properly captured. It attests to the artifact, not to reality. That is why Inficy only allows certification of eligible canonical bundles produced by runtime-observed integrations.

Agent-reported sessions, including ChatGPT connector checkpoints, are never certifiable. They are self-reported by the assistant, so no observation stands behind them.

What this does not establish

• It does not establish that an agent's output was correct or safe.
• It does not establish anything about steps the capture layer never observed.
• It does not establish the intent of the operator or the agent.
• It does not replace your own controls, review or approvals.

A practical sequence

Connect a supported runtime so executions are observed rather than reconstructed. Read the execution page to understand the steps. Export the bundle when you need it outside Inficy. Certify only the rare run that genuinely needs an independent third-party attestation.

Recording is the product. Certification is the exception, and it is optional.