Graduated autonomy is a deployment approach in which a system receives broader authority only after specified evidence supports the additional activity. It is a proposed control design, not a universal maturity scale or a consequence of improving a model’s benchmark score.
Capability and permission have different objects
A model may become better at writing a database query. That improvement does not establish whether it should execute the query against production records. Execution adds questions about authorization, affected data, reversibility, and the ability to detect a mistake.
Research on extreme-risk evaluation distinguishes capabilities from the propensity to use capabilities harmfully. That distinction does not supply a ready-made permission ladder, but it prevents a capability result from answering every behavioral question.
The same reasoning applies to ordinary deployments. Producing a correct recommendation, preparing an action, and committing the action are different tasks with different consequences.
Promotion needs evidence about the next boundary
Consider an assistant currently allowed to draft purchase requests. Giving it authority to submit approved requests requires evidence about approval validation, amount limits, duplicate handling, and the receiving system. Allowing it to approve its own requests would cross another boundary.
These examples are design deductions, not claims that a particular benchmark has validated those controls. The evaluation should name the permission being added and the failure conditions that matter for it.
Authority can also move downward
A model change, new tool, expanded customer population, or altered external process can invalidate assumptions behind an earlier approval. A promotion record should therefore include conditions for reassessment or reduced authority.
Retaining an advisory mode can allow useful work while consequential actions return to human review. That is a change in the system’s permitted behavior, not a declaration that every capability has disappeared.
The relevant progression is from evidence to a bounded permission. A higher aggregate score can contribute to the evidence, but it cannot identify the action, its consequences, or the authority that should be delegated.