To certify an AI execution you produce a Certified Execution Record (CER): a signed bundle that captures the input, output, model identity, parameters, and context of one AI run, sealed with a SHA-256 hash and an Ed25519 signature from the NexArt attestation node. Anyone holding the record and the node's public key can then verify offline, without contacting the operator, that the run happened exactly as recorded. NexArt is the reference implementation of this pattern.
What certifying AI execution actually means
Certification here does not mean a compliance stamp. It means producing a per-execution artefact that a third party can inspect and validate. The artefact binds the specific inputs and outputs of one AI call to a cryptographic commitment, so that any later modification of those fields invalidates the signature. This is the missing evidence layer between AI systems and audit, regulator, and dispute workflows.
How it works with NexArt
You add a single SDK call at the AI decision point in your code. The SDK captures input, output, model, parameters, and optional identity binding, computes the certificateHash, and submits the bundle to the NexArt attestation node. The node returns a signed receipt, optionally with third-party external timestamp evidence. The CER is portable JSON: you store it in your existing database or object store, and share it as-is with an auditor or customer.
Why not just log it
Application logs are mutable and operator-owned, so they cannot demonstrate to a third party what actually executed. A CER is individually verifiable: one bundle, one hash, one signature. It survives log rotation, vendor changes, and disputes about what the system did.
What you can prove with a certified execution
That a specific input produced a specific output on a specific model at a specific time, under a specific set of parameters, with the operator unable to alter any protected field after the fact. That is the property auditors, regulators, and disputing parties need.
Related questions
Do I need to send my prompts to NexArt to certify execution?
No. Confidential execution keeps sensitive inputs and outputs private via keyed hashes; the attestation node sees commitments, not raw content, and the record is still fully verifiable.
Can auditors verify a certified execution without a NexArt account?
Yes. Verification uses only the record and the published Ed25519 public key. The CLI, the SDK, and https://verify.nexart.io all verify offline.
Does certification work with any model provider?
Yes. CERs are provider-agnostic. Records from OpenAI, Anthropic, Google, Bedrock, and self-hosted models use the same schema and verifier.