Tamper-Evident AI Logs

    Standard logs record events. Tamper-evident logs make modification detectable by anyone.

    A tamper-evident AI log binds each execution to a SHA-256 commitment over canonicalised fields and an Ed25519 signature from an independent attestation node. Any later change to inputs, parameters, outputs, or metadata invalidates the certificate hash. This is what regulators, auditors, and customers mean when they ask for tamper-proof AI logs, tamper-evident is the accurate technical term.

    Side-by-side comparison

    DimensionStandard AI logsTamper-evident AI logs (NexArt CER)
    StorageApplication or observability databasePortable JSON artifact, verifiable offline
    IntegrityNone; the operator can modify entriesSHA-256 canonical hash over protected fields
    Third-party signatureNoneEd25519 signature from an independent attestation node
    Detection of editsNot detectable without external comparisonAny modification breaks the certificate hash immediately
    Retention controlControlled by the operatorControlled by whoever holds the record
    Verifiable by outsidersNoYes, with the published public key
    Regulator-readyNo, treated as operator narrativeYes, cryptographic evidence independent of the operator

    Why 'tamper-proof' is the wrong word

    No log is truly tamper-proof; any record can be deleted from a database. What matters is tamper-evidence: can a third party detect that a record has been altered without trusting the operator? A NexArt Certified Execution Record answers yes. The certificate hash covers the exact bytes of every protected field, and the attestation node's signature commits the hash to an external key.

    This distinction is what the EU AI Act Article 12 record-keeping obligations require in practice, evidence that survives the operator and can be checked by an outside party.

    Frequently asked

    Is a tamper-evident log the same as an immutable log?

    No. An immutable log prevents modification through storage guarantees. A tamper-evident log allows deletion but makes any content modification cryptographically detectable. Tamper-evidence is a stronger practical property because it does not depend on trusting a specific storage layer.

    Do I need to store raw AI inputs and outputs to have a tamper-evident log?

    No. In confidential mode, sensitive fields are replaced with SHA-256 commitments before permanent storage in the confidential proof record. The commitment is what gets signed, so the record proves what executed without exposing the underlying data.

    Add verifiable execution to your AI stack

    Produce Certified Execution Records anyone can verify without access to your infrastructure.

    Related