Humans are assigned purpose, priority and escalation judgement rather than simply obeying the model.
Proposed mechanism: Keep consequential choices with accountable people.
Conditions: Users can inspect and challenge model outputs.
Limit: No usability or oversight-performance evidence provided.