Risk tiering

The practice of scoring every distinct task type an agent performs against a fixed set of questions, severity of impact, reversibility, and feasibility of human oversight in the Dayos method, and grouping task types with similar scores into named tiers, each carrying a defined reasoning strategy and autonomy decision.

Sorting an agent's task types by severity of impact and reversibility of action, then assigning each tier a matching level of autonomy (full automation, automation with mandatory human sign-off, or no autonomous execution at all). Dayos's published case study is the framework's worked model of this practice.

Sorting work into levels (for example low, standard, high) by the consequence of an error, and attaching the appropriate review to each tier, with explicit triggers that let members self-sort.

Sorting AI systems into risk categories (for example, the EU's unacceptable / high-risk / limited-risk / minimal-risk) so obligations scale with risk. The first question of the regime reader.