What is AI execution integrity?§
AI execution integrity is the ability to produce a tamper-evident, verifiable record of a specific AI execution. It establishes that a recorded decision event occurred as captured and that its protected fields have not been altered since capture. Integrity is a property of the recorded artifact, not of the underlying model behavior.
What is an execution artifact?§
An execution artifact is a structured, machine-verifiable representation of an AI decision event. It contains the fields required to identify the execution, bind it to its inputs and outputs, and support deterministic verification. Artifacts are designed to be portable across systems and verifiable independently of the system that produced them.
How is this different from logs or observability?§
Logs and observability data are operational signals intended for monitoring, debugging, and aggregation. They are typically mutable, lossy, and scoped to a single operator's infrastructure. Execution artifacts are designed for deterministic verification and audit: their protected fields are stable, independently verifiable, and intended to outlive the originating system.
Does AIEF require deterministic AI outputs?§
No. AIEF does not require reproducible model outputs or deterministic inference. It requires integrity of the recorded artifact: that the captured representation of the execution can be verified as unmodified. Non-deterministic models can produce AIEF-conformant artifacts.
Does verification prove the model was correct?§
No. Verification proves integrity, not correctness, fairness, or appropriateness. A PASS result indicates that the artifact has not been tampered with and conforms to the specification. It does not assess whether the underlying decision was accurate, ethical, or fit for purpose.
What does "tamper-evident" mean in practice?§
Tamper-evident means that any modification of protected fields is detectable through deterministic verification. Verification does not prevent modification; it ensures that modification cannot occur silently. Any change to a protected field causes verification to FAIL.
What is deterministic verification?§
Deterministic verification is a verification procedure that produces consistent PASS or FAIL results for the same artifact, using defined rules and stable serialization. Independent verifiers applying the specification to the same artifact must reach the same conclusion. Determinism is a property of the verification procedure, not of the model.
What is the minimum implementation to get started?§
The minimum entry point is Level 2: an execution artifact, tamper-evidence over protected fields, and deterministic verification. Level 2 establishes the integrity guarantees that distinguish AIEF artifacts from operational logs.
When is higher conformance (Level 3 or 4) needed?§
Higher conformance levels apply when artifacts must support external audit, multi-step workflow traceability, or dependency relationships between executions. Level 3 introduces independent validation and richer execution context. Level 4 addresses independent attestation and trusted timestamping for use cases requiring external trust anchors.
Who should implement AIEF?§
AIEF is intended for systems whose AI executions have external impact, audit requirements, or long-term traceability needs. This includes regulated domains, decisions affecting third parties, and workflows where evidence of a specific execution may be required after the originating system is unavailable.
Related questions§
This page helps answer:
- How to verify AI output integrity
- What is a tamper-evident AI record
- How to audit AI decisions
- What replaces logs for AI governance
- How to create verifiable AI execution records
Where to go next§
New to AIEF?
Looking for implementations?