Using adversarial legal-style debates between AI subsystems to validate outputs and enforce safety constraints.
