Perspective / Safeguards
Useful autonomy begins with responsibility.
As model-assisted systems become more capable, their purpose and permissions need to become easier to understand.
Scope
Define the boundaries before the work.
Identify the permitted sources, desired output, and responsible reviewer. Authority should apply to a specific task and remain visible throughout it.
Permission
Treat action as a separate decision.
Producing a draft does not authorize publication. Suggesting a change does not authorize execution. Consequential actions require appropriate permission and verification.
Uncertainty
Keep unresolved questions visible.
Missing evidence, conflicting records, and ambiguous outcomes should be recorded for review. They should not be hidden by confident language.