An open protocol · vendor- and model-independent
What did the AI decide — and what does the evidence not cover?
AIREP — the AI Runtime Evidence Protocol — is a small, fixed record format. One signed, hash-chained, canonical-JSON record stands for exactly one runtime decision: what was decided, on what basis, the evidence behind it, and — distinctively — what that evidence does not cover.
A receipt for a single AI decision that an outsider can read and check, without access to the model that made it.
Status: Experimental — a proposed open format with one reference implementation. Not a ratified standard.
The gap it fills
Every system answers “why this decision?” differently — or not at all.
Telemetry conventions record what ran, not what was decided under whose authority. Model and system cards carry authority, but for a whole artefact, not a single decision. Supply-chain attestations sign builds and content, not per-decision dispositions. Governance frameworks mandate a logging capability while leaving the on-the-wire shape of a per-decision record undefined.
AIREP defines that missing object: a per-decision, governance-grade evidence record that anyone can read and check.
What a record carries
One decision, written down so it can be checked.
The decision
One verb from a closed set — release, block, defer, redact, escalate, kill — under a stated policy basis.
The evidence
Typed pointers to the supporting material, never the content itself, so a record stays small and shareable.
Scope-honesty
Every record declares, on the wire, both what its evidence covers and what it does not — so a checker knows what it cannot rely on.
Tamper-evidence
A SHA-256 hash chain binds each record to its position; altering or splicing a record breaks the chain.
Neutral core
All vendor-, framework-, or domain-specific content lives under a single profiles block. Delete it and the record still validates — a mechanical neutrality test.
Independently checkable
Two verifiers on different stacks (Python + Node) recompute every record and agree byte-for-byte. Replay is record-bound — it does not re-run the model.
Stated honestly
A candidate format and an epistemic discipline — not a certification.
AIREP is a proposed format with one reference implementation — the Phionyx runtime, which matures by conforming toward the format rather than defining it. The protocol carries no Phionyx physics, vocabulary, or dependency; that independence is the whole point. A second, independent producer is the largest open question, and the spec says so.
It does not make an AI truthful. It makes a single AI decision inspectable, bounded, and tamper-evident.
Archived & citable on Zenodo — DOI 10.5281/zenodo.20475136 (always resolves to the latest version).