The Need for Per-Decision Evidence

As AI systems are increasingly integrated into critical decision-making workflows, the lack of verifiable audit trails presents a significant governance challenge. Current monitoring often relies on opaque logs that are difficult to validate post-hoc. AIREP (AI Runtime Evidence Protocol) addresses this by formalizing a standard for generating cryptographic evidence at the point of decision, ensuring that every AI output can be linked back to its specific inputs, model version, and execution context.

Core Protocol Mechanics

AIREP functions as a bridge between the AI inference engine and external governance verifiers. The protocol mandates that for every decision, the system must produce a cryptographically signed evidence package. This package includes:

• Input Provenance: A hash of the exact prompt or data provided to the model.
• Model Identity: A unique identifier for the model version and configuration used.
• Execution Context: Metadata regarding the runtime environment and any applied guardrails.
• Cryptographic Proof: A signature that ensures the integrity of the evidence, allowing third-party verifiers to confirm that the decision was made by the claimed model under the specified conditions.

Implementation and Conformance

The authors provide a reference implementation and a two-verifier conformance kit to demonstrate how organizations can integrate AIREP into existing pipelines. By standardizing the evidence format, the protocol allows for interoperable auditing, where multiple independent verifiers can check the same decision logs without needing access to the underlying model weights or proprietary infrastructure. This approach balances the need for rigorous oversight with the practical constraints of production AI deployment, moving governance from a reactive "black box" audit to a proactive, verifiable runtime process.