You need a record of the execution whose integrity does not depend on you. A Certified Execution Record binds the input, output, model version, parameters, and context of a single AI execution into a canonically hashed artifact, signed by the NexArt attestation node and timestamped through a public RFC 3161 authority. Any party, including one that distrusts you, can verify offline that the record is the one produced at that time and has not been altered since.
Why application logs are weak under challenge
Logs, traces, and database rows sit inside infrastructure the operator controls. They can be edited, regenerated, sampled, or rotated, and nothing in the artifact itself distinguishes an original entry from a later one. A challenger does not have to prove tampering occurred; it is enough to observe that the record's custodian had the ability and the motive to change it.
Model drift compounds the problem. Re-running the prompt today does not reproduce what the model returned last March, so replay is not a substitute for a record.
What a Certified Execution Record establishes
- Which model, version, and parameters were used for that execution.
- The exact input and output, as raw values or as keyed commitments.
- An optional identity binding tying the execution to a case, account, or subject.
- That the record has not changed: any edit to a protected field breaks the certificate hash.
- When the record existed, anchored by an RFC 3161 timestamp issued via DigiCert.
Proving it without disclosing the data
Disputes often require proof to a party that has no right to see the underlying material. Confidential execution is the default: inputs, outputs, and identity fields are stored as keyed cryptographic commitments. The counterparty can verify that a specific value matches the commitment when you choose to reveal it, and can verify the record's integrity and timestamp even when you do not.
Scope and limits
Verification proves integrity, not correctness. A Certified Execution Record shows what the system executed. It does not show that the decision was correct, fair, or lawful, and NexArt does not assert legal admissibility. Evidentiary weight depends on the jurisdiction, the contractual context, and the organisation's own procedures. The record is designed so that its integrity can be established without relying on the operator or on NexArt.
Where teams start
Certify the executions most likely to be contested: adverse decisions, eligibility and pricing outcomes, content moderation actions, automated communications, and agent actions with external effect. Multi-step flows are grouped into a Project Bundle with a single Project Hash so ordering across steps is preserved.
Related questions
Is a Certified Execution Record admissible in court?
NexArt does not assert legal admissibility. Admissibility and weight depend on the jurisdiction, the contractual context, and the operator's own procedures. The record is built so that its integrity is independently verifiable, which is the technical property those assessments usually turn on.
Can we produce evidence after the fact?
No. A record must be produced at execution time. Certification is a call at the AI decision boundary; there is no way to certify an execution retroactively, which is precisely why the record carries weight.
What if the dispute involves personal data we cannot share?
Confidential execution stores keyed commitments instead of raw values, so integrity and timing can be verified without disclosure. Specific values can be revealed selectively later and checked against the commitment.
Does this apply to AI agents, not just single model calls?
Yes. Each agent step, tool call, and decision can emit its own record, and related steps are grouped into a Project Bundle so the sequence itself is covered by cross-step integrity.