What verification actually proves

    Integrity, not correctness. Provenance, not truth. This distinction is the whole point.

    A verified NexArt record is a precise, narrow claim. Being precise about what it proves, and what it does not, is what makes it credible. Overstating the claim would make the system worthless to auditors and regulators.

    What a verified CER proves

    • Integrity. The inputs, outputs, parameters, and context in the record have not been altered since the record was sealed.
    • Provenance. The record was produced at a specific time, attested by the NexArt attestation node under a specific signing key, and identified by a specific certificate hash. Execution ID, when present, is an application-level identifier for locating records; certificate hash is the exact CER identity.
    • Binding. The inputs and outputs are bound together. You cannot silently swap one for the other without invalidating the hash.
    • Independence. Anyone can recompute the hash and verify the signature without trusting NexArt, the operator, or the original system.

    What a verified CER does not prove

    • Correctness. It does not prove the model's answer was right, accurate, or useful.
    • Truth of content. If the model hallucinated, the hallucination is faithfully recorded. Verification proves the output was produced, not that the output is true.
    • Provider identity. It does not prove which underlying model provider was actually called, only what the operator declared at the execution boundary.
    • Policy compliance. It does not, by itself, prove a decision followed your internal rules. It gives you the evidence to check, but the policy check is separate.
    • Intent. It does not prove why a system produced what it produced.

    Why a narrow claim is still valuable

    Most disputes about AI systems are not about whether an answer was correct. They are about whether an answer was actually produced as described, with the inputs described, at the time described, or whether the record was edited after the fact. Integrity and provenance are the exact questions regulators, auditors, insurers, and counterparties care about.

    Correctness is a separate problem, and it is your problem. Verification gives you the stable ground on which any correctness, policy, or quality analysis can stand.

    Two common confusions

    Common claim: "A verified record means the AI did the right thing."

    Reality: No. It means the record of what the AI did has not been tampered with. Whether what it did was right is a separate judgment.

    Common claim: "If I do not trust the output, verification is pointless."

    Reality: Verification is most useful precisely when you do not trust a party. It lets you check what they claim happened without trusting them to tell the truth about it.

    The practical summary

    A NexArt verification answers: Did this execution, with these inputs and outputs, actually happen as recorded, or has the record been altered? That is the question. Anything broader is outside the scope of the protocol, and we will not claim otherwise.