Protocol proposed to record per-decision evidence for AI runtime governance

A new protocol, AIREP, is proposed to standardize how automated AI runtimes record governance decisions such as releasing, blocking, deferring, redacting, or escalating individual outputs. Each decision is captured as a single signed object that can be checked offline by any party, independent of the runtime that produced it.

Each AIREP record includes a closed set of verbs to describe the decision, a stated policy basis, and cryptographic references (by hash) to the input, output, and evidence—rather than embedding the actual values. This design limits exposure of sensitive data while enabling verification.

Records are organized into a SHA-256 hash chain, binding each record to its position in the sequence. This structure makes tampering or omissions detectable through recomputation, supporting integrity and auditability.

Vendor-, model-, and domain-specific content is confined to an optional namespace, and a mechanical neutrality test is applied to ensure the shared format remains free of such content. This separation aims to preserve interoperability and neutrality across implementations.

The protocol includes a reference implementation and a two-language conformance kit to support adoption. The authors also discuss implementation challenges, including alignment of canonical forms across implementations, freshness witnesses, and multi-runtime chains.

The format is offered for adoption by any AI runtime that records governance decisions, positioning AIREP as a potential standard for transparent, verifiable AI runtime governance.