A facilities agent can update an incident record. It can also close routine maintenance notices at its assigned site. Those two permissions sound compatible until an unexplained sensor alert appears. In the paper’s hypothetical application, marking that alert resolved can remove it from the unresolved-incident display and send an all-clear message. The ability to change a status field has become a decision with consequences for other people.

My new preprint, Bounded Agency: Ontology-Driven Authority for Delegated Intelligence, examines how to define that decision boundary and keep changes in meaning from silently moving it. It is available on Zenodo as version 0.1.0-preprint, under Creative Commons Attribution 4.0.

What has actually been delegated?

The paper separates four questions. What can the system technically do? Which operations may it invoke? Which situations and outcomes has it been entrusted to decide? Who may approve a new model as the basis for those decisions?

A mandate answers the third question: it specifies the work and judgment delegated by a person or organization. A support agent might have authority to issue a small courtesy credit while lacking authority to cancel a contract. A facilities agent might close an independently verified routine notice while lacking authority to declare an unexplained situation safe. Access to the same customer or incident record does not settle those distinctions.

Existing authorization systems can express sophisticated business conditions. The issue is whether the deployment actually represents the intended decision boundary, its evidence requirements, and its downstream effects. Bounded agency names the scope of independent judgment that those controls must enforce.

A category change can change a decision

An ontology describes the concepts and relationships a system uses to interpret its domain. In the facilities example, a policy allows closure of a RoutineNotice. The current model treats service notices as routine, but does not place an unexplained SensorAnomaly in that category.

Now suppose an improvement process proposes a new relationship: every sensor anomaly is a kind of routine notice. Under that candidate model, the unchanged policy can permit a closure that it previously refused. Neither the agent’s identity nor the policy text has changed. The meaning of a term used by the policy has.

The paper deliberately keeps the example narrow. Its observation checks establish provenance and completeness, not clearance of a possible hazard. If the policy separately required independently verified safety clearance, a category edit would leave that requirement in place. The example shows why the chosen category boundary matters; it does not demonstrate that any alert is safe to close.

Compare the change before allowing its use

The proposed review asks which decisions become eligible under the new model when the agent, policy, evidence, and relevant context are held fixed. It also asks which previously eligible decisions disappear. A concrete before-and-after case gives reviewers something more useful than an assurance that the ontology remains internally consistent.

The paper calls this authority-relative semantic-change admission. It means evaluating a change in meaning against a particular delegation before accepting that change for operational use. The comparison can run without authorizing any real action. A candidate passing that hypothetical evaluation still needs an active grant that admits the approved model and policy.

Two approval questions remain distinct. A domain steward can decide that a new classification represents the domain accurately. The person or institutional role entitled to delegate the work must decide whether the resulting decision scope is acceptable. A better description can be adopted while autonomous resolution of the newly described cases remains withheld.

The operational grant identifies the approved versions. Before executing an action, the enforcement path checks that the grant remains current and that the actual operation matches the authorized outcome. An agent’s recommendation cannot be treated as harmless advice if another system automatically commits it as a final decision.

Facts need an admission rule too

The ontology is only one route to a changed decision. An agent could leave the hierarchy untouched and write “this alert is routine” into the situation store. If the policy relies on that assertion, the result can change just as surely.

Recording an observation and attesting that it meets a decision condition are different assignments. The approved process must say who can supply that classification and what evidence supports it. Ordinary new facts should still change decisions when the mandate expects them to; review should focus on their admissibility and authority impact.

How this relates to the lifecycle paper

The Orchestrator lifecycle paper describes the path through operating, investigating, developing, and releasing changes. Bounded Agency concentrates on a specific obligation within that path: identify how a proposed semantic change affects delegated judgment and route that effect to the role entitled to approve it.

That connects to all three books. The professional needs to recognize when a situation exceeds the assignment. The engineer needs to represent the decision and mediate its effects. The organization needs to name who may revise the mandate and who remains answerable for the outcome.

This is a conceptual preprint and has not been peer reviewed. It builds on established semantic-policy, ontology-evolution, and authorization research. Its executable illustration compares four fixed proposals in a small taxonomy; it reports no production deployment or measured improvement in agent reliability. The research question is whether the explicit delegation-relative review adds useful detection or reduces review burden beyond strong existing controls.

For one workflow you already automate, choose a category that appears in its decision policy. Identify who can change that category, who can attest membership, and which decisions would change as a result. Then name the role that may approve that change in authority. Those answers make the delegated judgment inspectable.

Read Bounded Agency on Zenodo, including its research supplement, or inspect the public manuscript and executable illustration. Future versions will be connected through the all-versions DOI.