The standard

What ODES actually is

A vendor-neutral, portable decision-evidence record format for one narrow, critical job: letting decisions carry structured governance metadata across a boundary — without exposing raw evidence, requiring shared infrastructure, or forcing anyone onto a proprietary trust network.

The core object

The decision evidence record

The record is the whole point of ODES. It is not the raw data, not the model, not the full audit system, and not the workflow that produced the decision. It is a portable governance envelope for a decision: a structured evidence container that answers one relying-party question:

What was decided, by whom or by what, under whose authority, with what human disposition and machine role, against what evidence and policy basis, under what model state, and with what freshness?

Crucially, the record does not have to expose the evidence itself. It can carry references, hashes, attestations, verification material, policy identifiers, risk coordinates, freshness, revocation, and consumption conditions — so confidential material stays with the issuer while the relying party still gets structured evidence it can evaluate under its own rules.

Field vocabulary

What a record captures

A plain-English view of the governance coordinates a decision should carry when it crosses a boundary. The full field summary and schema profiles live in the repository.

Authority

The basis on which the decision could be made, and by whom.

Machine role

Whether AI retrieved, recommended, executed, or escalated.

Human disposition

Whether a human reviewed, modified, overrode, or simply accepted.

Model state

Version, tool access, and runtime conditions that shaped the output.

Evidence & policy basis

References, hashes, and policy identifiers — not the raw evidence.

Freshness

Expiration, revocation, supersession, and consumption conditions.

Verification

Structure is not reliance

One of the most important distinctions in ODES: a well-formed record is not the same as a trusted one. Each layer means something different.

Layer
What it tells you
Schema validity
The record is well-formed against the ODES schema.
Profile conformance
It meets the rules of the specific profile it claims to follow.
Verifier acceptance
A verifier's checks — signatures, freshness, references — pass.
Relying-party reliance
The receiving party decides, under its own rules, to actually rely on it.

Where it fits

Complementary, not winner-take-all

ODES sits between the tools you already have. It's the portable decision-evidence layer that lets them interoperate across a boundary.

Tool
What it answers
Governance / GRC
ISO 42001, NIST AI RMF
Does the organization have controls? (Mostly internal.)
AIBOM
What is this AI system made of?
Provenance
What is the lineage and trace of the data or model?
ODES
What was this specific decision, and what evidence can another party evaluate before deciding whether to rely?

Related work

Boundary blindness

ODES is informed by adjacent work on provenance, AI governance, model documentation, and decision traceability. Alexander D. Barrett’s July 2026 paper, Boundary Blindness Under Artificial Intelligence, is especially important related work because it frames the problem as a cross-boundary decision-evidence gap that AI accelerates rather than creates. ODES should be read as one candidate standards-layer response to that diagnosis. No endorsement, affiliation, sponsorship, or relationship is implied.

Status

An open, early draft

v0.2 · discussion draft

ODES is a candidate open standard, not a finished specification. The strong claim is narrow: ODES defines the portable decision-evidence record as the interoperability object for AI-influenced decisions that cross boundaries. It addresses the part that governance frameworks, model documentation, provenance systems, and audit trails often leave under-specified.

It is developed in the open and meant to be shaped by the people who would issue, verify, and rely on these records. The discussion draft, schema profiles, examples, and field summary are maintained in the public repository.

Read the discussion draft ↗