Machine-coached policy revision lets AI regulators explain and fix their own decisions
A new controller-level contestability layer uses defeasible rules with explicit priorities to enable in-loop policy improvement.
Roberto Garrone's paper proposes a machine-coached policy-revision layer for adaptive agent-based regulatory simulation. It represents policy decisions as defeasible rules with explicit conflicts and priorities, enabling explanations and diagnostic failures to trigger rule additions, removals, or priority changes. In a controlled emissions-regulation ABM experiment, the predefined coaching template adds a relaxation rule to the symbolic controller, reducing over-conservatism recurrence under held-out seeds while preserving violation, overshoot, and volatility guardrails. The approach makes controller-level contestability operational: policies can be explained, challenged, revised, and re-evaluated in held-out simulation runs.
- Represents policy decisions as defeasible rules with explicit conflicts and priority ordering, enabling in-simulation explanation and revision.
- Diagnostic failures are automatically translated into rule additions, removals, or priority changes via a coaching template.
- In a controlled emissions-regulation ABM, the layer cut over-conservatism recurrence by 60% while preserving violation, overshoot, and volatility guardrails.
Why It Matters
Machine-coached contestability makes AI-driven regulation transparent and adaptive, building trust in autonomous policy systems.