You use confidential execution: sensitive inputs and outputs are replaced with keyed cryptographic commitments (HMAC-SHA-256) inside your environment before the record leaves it. The third party receives a Certified Execution Record they can verify end to end, sees the model identity, parameters, and timestamps, and confirms that the commitments were sealed intact, without ever seeing the raw prompt, response, or user data. NexArt is designed around this constraint: the attestation node sees commitments, not content, and verification works offline.
The problem: you owe evidence but not data
Regulators, customers, auditors, and disputing counterparties increasingly ask for proof of what an AI system did. But the underlying prompts and outputs often contain regulated data, trade secrets, PII, or confidential business content that you legally or contractually cannot share. Handing over raw logs is not an option. Refusing to hand over anything is also not an option.
The solution: confidential execution
Confidential execution separates the evidence from the content. Sensitive fields are replaced with keyed cryptographic commitments before it is written into the confidential proof record. The commitment binds the value: you can reveal it later to a specific party under NDA if you choose, and they can verify it matches. Until then, only the commitment travels. The rest of the record (model, parameters, timestamps, node signature) is fully visible and verifies normally.
What a third party can verify without seeing the data
That a specific execution happened at a specific time on a specific model with a specific parameter set. That the commitments over the sensitive fields were sealed at that time and have not changed since. That the attestation node's signature is valid. That the record has not been altered. That a Project Bundle groups a coherent multi-step trace. That any commitment you later reveal matches the one that was sealed.
What this unlocks
Selective disclosure to auditors, regulators, or counterparties under NDA. Public verification of execution integrity without any content leak. Compliance evidence for EU AI Act Article 12 in industries where content cannot leave the environment. Vendor assessments where a customer wants proof that your AI ran as specified, without you exposing prompts or outputs.
Related questions
Does the attestation node see my prompts or outputs?
No. In confidential execution mode, sensitive fields are replaced with keyed HMAC-SHA-256 commitments in your environment. The node signs the commitment, not the content.
Can I reveal one field to one party later without exposing everything?
Yes. Selective disclosure works per field. You hand the counterparty the raw value plus the commitment key for that field; they recompute the commitment and confirm it matches the sealed one.
Does this weaken the verification?
No. The signature and hash chain are identical to non-confidential execution. Third parties get the same integrity guarantee; they simply cannot read the content of redacted fields until you choose to reveal them.