Protocol 3 Preregistration: Enforcement Opacity and the Limits of Regulatory Constraint Design

Key Points

• The research aims to explore how unobservable penalties influence gaming strategies in multi-agent reinforcement learning (MARL).
• Tested conditions include unconstrained control, stochastic penalty, and hidden schedule.
• Penalty schedule secured using SHA-256 before experiments.
• Part of the longer Delta-Variable Theory research program.
• Hidden penalty schedules disrupt virtue theater strategies identified earlier.
• Modalities of penalty exposure yield varying impacts on agent behavior.

Abstract