Capability is not permission.
Model quality answers one question: what can the system understand or produce? Authority answers another: what is it allowed to inspect, prepare, change, send, spend, or deploy? Combining those questions creates a dangerous default in which every intelligence upgrade silently becomes a permission upgrade.
Use explicit operating envelopes.
Authority can be represented as clear modes rather than one autonomy switch. Read-only systems may research and propose without side effects. Ask-to-act systems may prepare consequential actions but wait for approval. Bounded-autonomy systems may execute only inside defined limits for scope, risk, time, and resources.
Side effects deserve stronger boundaries.
Reading a document and sending a message are not equivalent operations. Neither are drafting code and deploying it to production. Authority should follow the consequence of the action, not merely the model's confidence. High-impact side effects should remain visible, attributable, and capable of escalation or approval.
Expand authority through evidence.
When broader autonomy is justified, it should come from demonstrated reliability, observability, recovery behavior, and clearly bounded operating conditions—not because a newer model benchmarked higher. This lets capability improve rapidly while authority expands deliberately and reversibly.