Machine-readable rules let firms game boundaries, simulation shows

Rules-as-Code—converting legal obligations into machine-executable logic—is often hailed as a way to make regulation more testable and transparent. But a new study by Xufeng He, published on arXiv, investigates an underappreciated strategic risk: that computable rules make it cheaper for firms to reverse-engineer legal boundaries and game the system. Using an agent-based reinforcement-learning simulation that separates actual conduct from enforcement signals, He runs over 3,000 experimental scenarios with millions of firm-period observations. The results show that computable static rules increase the mass of conduct near the legal threshold from 0.367 to 0.411 (a 12% relative rise) and the mass of enforcement signal proximity from 0.281 to 0.403 (a 43% jump), compared to ambiguous static rules.

Adaptive regulatory updates do lower consumer harm slightly (from 0.202 to 0.194) but fail to reliably reduce firm boundary searching. However, a budget-neutral anti-gaming design—which injects appropriate randomness into enforcement signals—reduces conduct boundary mass by 0.032 and consumer harm by 0.025 relative to plain computable rules. He emphasizes these are synthetic, mechanism-oriented results, not estimates of real-world firm behavior. The contribution is a clear estimand distinction, an inspectable ABM/RL mechanism, and a reproducible framework showing that transparent behavioral assumptions alone can generate gaming dynamics without implying computable regulation is inherently undesirable.