Skip to content

Agentic AI Needs Consequential Evidence

The United Nations Co-Chairs' Summary of the inaugural Global Dialogue on Artificial Intelligence Governance describes an international policy conversation increasingly concerned with verifiability, traceability, interoperability, evidence, accountability, and agentic AI.

The report does not discuss or endorse OpenETR. It is also a summary of themes observed by the Co-Chairs, not a negotiated outcome or consensus position. It nevertheless provides an important external signal: as AI systems become more capable of taking consequential actions, governance will need evidence that can be evaluated beyond the application in which an action occurred.

OpenETR offers a useful architectural response:

AI may infer what a digital artifact means. Consequential action should rest on independently verifiable evidence, identified rules, and an explicit recognition decision.

Policy Context

The UN summary identifies several themes that are directly relevant to this problem.

First, participants described interoperability, rather than complete harmonization, as a realistic objective for connecting different national and regional governance approaches. The report points to common technical foundations, shared definitions, comparable classifications, open benchmarks, and mutual recognition of testing and audit protocols.

Second, it presents verifiability as a basis for trust. Participants called for independent testing, disclosure across the AI value chain, incident reporting, and exchanges of evidence.

Third, it identifies increasingly autonomous and agentic AI as a particular accountability challenge. The proposed elements of accountability include transparency and traceability, documentation of agent chains, an identifiable and empowered human decision-maker, clear liability, remedies, and audit capacity.

Fourth, the report calls for practical and interoperable tools that can move AI governance from principles toward implementation.

These themes do not prescribe a technical architecture. They do, however, expose the need for one.

The Hard Problem Has Moved

The article The Machine Can Read the Document. But Should It Believe It? provides the complementary technical insight.

AI is rapidly reducing the cost of interpreting unfamiliar or weakly structured documents. An agent can often understand an invoice, warehouse receipt, purchase order, certificate, image, or handwritten note without every participant first adopting the same application or schema.

But a fraudulent invoice may be just as intelligible as a genuine one. Comprehension does not establish:

  • who issued the artifact;
  • whether the issuer or another actor was authorized;
  • which consequential actions occurred;
  • whether an authorization was revoked or superseded;
  • what current state follows from the evidence; or
  • whether another party should act on that state.

The policy problem therefore moves from making artifacts machine-readable to establishing reliable grounds for action.

From Agent Logs To End-Verifiable Evidence

Documenting an agent chain is useful, but an application log remains an assertion made and controlled by that application. A relying party may be unable to verify the log without privileged access to the originating system, its database, and its interpretation of current state.

For consequential actions, stronger documentation should be:

  • attributable to identifiable signing keys;
  • bound to the exact Digital Artifact or object concerned;
  • cryptographically linked to relevant prior evidence;
  • independently retrievable where disclosure policy permits;
  • evaluated under identified and versioned rules; and
  • capable of producing a reproducible state assessment outside the originating application.

OpenETR calls the resulting evidence structure a Digital Controllable Record (DCR). The DCR does not make every signed assertion true or effective. It provides portable evidence from which a verifier can determine what follows under a selected rule book.

AI interpretation
  -> proposes what the artifact means

Digital Artifact
  -> identifies the exact content by digest

DCR evidence
  -> records attributable consequential statements and relationships

Defined rules
  -> validate the evidence and derive Consequential State

Recognition
  -> determines whether a relying party accepts that result

Effect
  -> determines what action or external consequence follows

This division lets AI perform the interpretation at which it is increasingly capable without allowing inference alone to manufacture authority or current state.

A Consequential Evidence Layer For Agentic Systems

Consider an AI agent presented with a purchase order for $47,500.

The agent may accurately infer the supplier, amount, goods, delivery terms, and requested payment. Those are semantic conclusions. Before approving payment, the relevant system may need to determine:

  1. Did the identified organization issue this exact purchase order?
  2. Which key signed the approval evidence?
  3. What identity, role, or mandate is associated with that key?
  4. Was the approval within the applicable monetary and temporal limits?
  5. Was the authority later revoked or the purchase order superseded?
  6. What current authorization state follows under the selected rules?
  7. Does the paying organization recognize that evidence and state for this transaction?

OpenETR can provide the object-specific evidence and state-derivation layer. It does not perform KYC, assign organizational responsibility, determine legal liability, or compel payment. Those remain responsibilities of identity systems, host applications, governance frameworks, contracts, law, and relying parties.

Actor-Neutral Evidence, Human-Centred Accountability

The UN report emphasizes meaningful human oversight, an empowered human decision-maker, liability, remedies, and audit capacity. OpenETR does not replace any of these requirements.

At the protocol layer, a Key-Based Identifier identifies verification material. It does not establish whether the key was operated by a person, organization, service, device, or autonomous agent. This actor-neutral model is useful because real workflows frequently combine all of them: an agent proposes an action, a policy engine tests it, a human approves it, and a managed signing service applies the signature.

The integrating system and recognition policy must preserve the additional evidence needed to identify:

  • the accountable principal;
  • the agent, operator, or service involved;
  • the authority or mandate under which it acted;
  • the human approval or override path, where required;
  • applicable limits, expiry, suspension, and revocation;
  • the responsible provider or deployer; and
  • available review, appeal, and remedy processes.

Cryptographic attribution strengthens accountability evidence. It does not decide where accountability or liability ultimately rests.

Interoperability Without Harmonized Recognition

The report's emphasis on interoperability among differing governance regimes aligns closely with OpenETR's recognition boundary.

Participants do not need one shared application, database, identity provider, or legal conclusion. They can instead agree on enough common evidence and verification conventions to evaluate the same signed records. Each jurisdiction, institution, or counterparty can then apply its own recognition policy.

shared artifact identity
  + portable attributable evidence
  + reproducible verification rules
  = interoperable assessment

local law, policy, contract, or institutional mandate
  = recognition and effect

The central policy proposition is:

Interoperable evidence does not require harmonized recognition.

This approach supports international cooperation while respecting differences in law, institutional authority, risk tolerance, and local context.

Policy Recommendations

  1. Distinguish interpretation from evidence. AI-generated semantic conclusions should not be treated as proof of issuance, authority, authorization, control, revocation, or current state.
  2. Require stronger evidence for consequential action. Assurance should increase when an agent can transfer value, exercise authority, alter rights, or trigger legal or operational consequences.
  3. Make agent-chain evidence object-specific. Accountability records should identify the exact artifact or transaction concerned and link to relevant prior evidence.
  4. Prefer portable evidence over application-only logs. Another authorized implementation should be able to verify the evidence without requiring the originating application to remain available.
  5. Identify the rule book. A state conclusion should disclose the rules, policy version, evidence scope, and verification time on which it depends.
  6. Report verification dimensions separately. Artifact integrity, signature validity, graph continuity, authorization, evidence sufficiency, actor recognition, and external effect should not collapse into one trusted Boolean.
  7. Preserve recognition plurality. Common technical evidence should support, not displace, jurisdictional and institutional authority.
  8. Retain human accountability mechanisms. End-verifiable evidence should complement human oversight, audit, appeal, remedy, and liability frameworks.

Implications For OpenETR

The report does not require a new OpenETR primitive or a special class of AI-authored event. The existing actor-neutral model already permits people, organizations, services, and agents to participate through signing keys while leaving identity, authority, accountability, and recognition to the applicable context.

The immediate OpenETR priorities are therefore to:

  • develop agentic-system examples using the existing DCR model;
  • improve representation of authorization, delegation, revocation, execution, and linked evidence;
  • expose verifier results in machine-readable, dimension-specific form;
  • document how external identity and agent-governance systems supply recognition evidence; and
  • demonstrate independent state reconstruction across two separately operated systems.

The idea can be summarized without diminishing either AI or institutional judgment:

Inference interprets. Evidence supports claims. Rules determine what follows. Recognition gives effect.

AI is making digital artifacts broadly intelligible. OpenETR is concerned with making consequential actions concerning those artifacts independently verifiable.