Control Evidence Object Model

A record model for what an AI agent is permitted to do, what it is expected to do, what it was observed doing, and the evidence for each.

Why it exists

Permissions for AI agents are well served by existing standards. Scopes, rich authorisation requests, token exchange and authorisation APIs all describe what an agent may do. Observed actions and the evidence for them have no common schema. So an organisation can hold a register, a policy, a committee and a risk assessment process and still be unable to state whether what its agents did matches what anyone intended.

This model records that layer. Twelve objects across four layers, from an obligation through to the evidence for it, holding three positions on every action rather than two: what the grants permit, what the accountable person expects, and what was observed. The divergence between the second and the third is where authority expands without anyone deciding that it should.

The documents

Specification	PDF · source
JSON Schema	ceom.schema.json, draft 2020-12
Field dictionary	ceom.field-dictionary.csv
Vocabularies	ceom.vocabularies.json
Validator	reference/validate.py
Citation	CITATION.cff

The twelve objects

Designed	1 Obligation · 2 Interpretation · 3 Control · 4 Owner · 5 Authority
Estate	6 Asset · 7 Dependency
Runtime	8 Action · 9 Exception · 10 Decision
Spanning	11 Evidence · 12 Permission

Conformance

Annex D distinguishes two kinds of failure that must never be mixed. A defect means the records are malformed, so nothing can be concluded from them. A finding means the records are well formed and say something about the subject. A null approver is not a broken record; it is a statement that nobody approved the grant.

Conformance is the absence of defects. It is not the absence of findings.

Identifiers

Terms are identified under this namespace, versioned:

Object	https://ceom.aiscale.ltd/0.9.10/object/obligation
Field	https://ceom.aiscale.ltd/0.9.10/field/obligation.source_ref
Term	https://ceom.aiscale.ltd/0.9.10/term/exception/transformation_defect

These are names. Only the schema resolves to a document today; the remaining term addresses are identifiers rather than endpoints, and will be served when there is a reason to.

Status, stated plainly

The twelve objects, the four layers, the evidence scale, the reference states and the gap scale are stable. Field-level detail may still move.

The model has been instantiated against two real agent assessments. The worked examples in the specification are constructed, deliberately, so that no reader receives a client estate along with the document.

What it still lacks is the measurement that matters: two assessors reaching the same determinations independently from the same artefacts.

Challenge it

The model was circulated for challenge rather than adoption. Open question 11 is the one that most needs other people: whether the Exception vocabulary survives contact with the parties who must map into it. Ten terms, each of which somebody's existing classification must fold into without loss.