What is usually missing is verification: evidence of what the model actually executed that a third party can check without trusting you. Governance produces documentation the operator writes about its own controls. Verification produces a tamper-evident record of each execution, sealed and independently signed, so its integrity does not depend on the operator's infrastructure or good faith.
Two different claims
A governance programme supports the claim "we had appropriate controls". Verification supports the claim "this specific output was produced by this model, with this context, and the record has not changed since". Auditors, regulators, customers, and opposing counsel usually arrive asking the second question, and most stacks can only answer the first.
Where the gap becomes expensive
- A customer disputes an automated decision and asks for the basis of it.
- A supervisory authority asks a deployer to reconstruct a specific high-risk run.
- An internal investigation needs to establish what an agent did, not what it should have done.
- A counterparty requires evidence of AI behaviour but is not permitted to see the underlying data.
In each case, application logs and observability traces are operator-owned and mutable. They can be edited, sampled, or rotated, so they carry little weight with a party that does not already trust the operator.
How execution evidence closes it
A Certified Execution Record captures the input, output, model context, parameters, and optional identity binding of a single execution, seals them with a canonical SHA-256 hash, and submits the result to the NexArt attestation node. The node returns an Ed25519 signature and an RFC 3161 timestamp issued via DigiCert's public timestamp authority. Any party can recompute the hash and check the signature offline against the published public key.
Confidential execution is the default: sensitive fields are stored as keyed commitments rather than raw values, so a record can be verified without exposing prompts, outputs, or personal data.
What this does not do
Verification proves integrity, not correctness. It does not show that a decision was fair, accurate, or lawful, and it does not make an AI system compliant with any regulation. NexArt covers execution and inference only. Governance regimes, conformity assessment, and compliance determinations remain with the organisation, its notified body, and its supervisory authority.
Related questions
Is this not what our AI governance platform already does?
Governance platforms hold policies, model inventories, risk assessments, and approvals. They record what the organisation intends and who signed off. They do not produce cryptographically sealed, independently verifiable records of individual executions, which is what a third party needs when an output is challenged.
Can our existing logs serve as verification?
Not on their own. Logs live inside operator-controlled systems and can be modified, sampled, or deleted. Verification requires evidence whose integrity can be checked without trusting the party that produced it.
Do we have to replace anything to close the gap?
No. A certification call is added at the AI decision boundary. Existing observability, logging, and governance systems stay in place and continue to serve their own purposes.