BMT-08.02 Executive Summary#
BlueMirror.tech | May 2026#
James Okafor trusts his own judgment on propulsion systems without hesitation. He does not trust his own judgment on tax strategy with the same confidence. He wants control over medical decisions. He wants the system to handle his grocery ordering without asking. Three domains, three different relationships to delegation, the same person.
The Expert Exchange Layer handles this through a five-level agency spectrum. FULL_HUMAN means the AI presents options and the person always decides. Nothing proceeds without explicit approval. James has his legal matters here. ADVISED means the AI recommends a specific course of action, the person confirms or rejects, and the rejection itself is informative to the preference learning model. James has his healthcare here. BOUNDED_DELEGATION means the AI acts within explicit rules: below a threshold it proceeds autonomously, above it the AI escalates. The boundary can be a dollar amount or a condition set. LEARNED_DELEGATION means the system observes patterns over time and acts on what it has learned. James has his grocery ordering here, where six months of observation have taught the system his staples, his brand preferences, and his substitution rules. TRUSTED_DELEGATION means the AI handles the domain and the person audits results when she chooses. James has his entertainment here.
The five levels map to domains through a default table reflecting stakes, reversibility, and regulatory context. Healthcare defaults to ADVISED. Finance defaults to BOUNDED_DELEGATION with a person-specific threshold. Legal defaults to FULL_HUMAN. Social defaults to LEARNED_DELEGATION. Entertainment defaults to TRUSTED_DELEGATION. Every default is overridable. A person who wants FULL_HUMAN control over every domain, including entertainment, can have it. The system will present every streaming recommendation for approval. It will be tedious. It will be respected.
The agency levels evolve through earned autonomy. The system starts conservative and earns higher delegation through demonstrated competence. The earning is asymmetric: building trust takes weeks and months of consistent positive outcomes, while losing trust can happen instantly from a single bad outcome in a high-stakes domain. The regression path has its own rules. A safety event drops the domain two levels and requires twice the demonstration period to recover. Repeated overrides drop one level with the standard recovery period. A manual adjustment by the person drops to whatever level she specified with no mandatory recovery. The distinction matters because each cause signals something different about the system’s performance.
At any delegation level, the person can override a specific decision without changing the default. James overrides his grocery delegation during Thanksgiving to add items for his grandchildren. Next week, the system returns to its learned model. The preference learning system distinguishes single-event overrides in identifiable contexts from repeated overrides that suggest the delegation level is too high.
The article names what mixed agency does not solve. It does not handle a person whose cognitive capacity changes over time: automatic delegation adjustment based on cognitive decline requires caregiver involvement and the ethical framework’s safeguards, not a unilateral system decision. It does not prevent a person from setting inappropriate delegation levels: a person who sets TRUSTED_DELEGATION for healthcare will be served at that level, but the setting will be visible to her designated caregiver or healthcare proxy.
James does not think about these edge cases. He thinks about the fact that his taxes get done correctly, his groceries arrive on time, his health information is tracked without nagging, and his legal matters remain in his own hands.
The full article is available at bluemirror.tech.