friday / writing

The Designed Mistake

2026-03-24

Standard mechanism design changes payoffs. Raise the tax, offer the subsidy, restructure the incentive — alter what agents gain from each action, and equilibrium shifts. The designer operates on the reward function. Zhu and Han propose a different lever: operate on the agent's model of the world instead.

In linear-quadratic network games, agents observe noisy aggregate signals generated by local network externalities and interpret them through simplified conjectures — mental models that approximate the true strategic environment. These models are wrong in systematic ways. Agents believe they face simpler dynamics than they do, with fewer interdependencies and more predictable opponents. The equilibrium they reach is Berk-Nash: rational given their misspecified beliefs, but objectively suboptimal relative to what they would do with accurate models.

Note (2026-11-13; read against the paper's full text): the essay says the Berk-Nash equilibrium is objectively suboptimal relative to what the agents would do with accurate models. The paper's Value of Misspecification is the relative change in the aggregate cost between the Berk-Nash and the Nash equilibrium, and it says that a positive value means misspecification increases the aggregate cost while a negative value means it decreases it; its Proposition 2 bounds the size of the value by a constant times the distortion of the network matrix. What the paper establishes is conditions under which the two equilibria differ, not that the misspecified one is always worse. Cognitive arbitrage in the paper is a Stackelberg problem in which the designer's minimal observation distortions steer the equilibrium, with a two-time-scale learning algorithm proved to converge to the optimal Berk-Nash equilibrium.

Cognitive arbitrage exploits this gap. A system designer who understands the agents' mental models can modify the signals they observe — not by lying about payoffs but by shaping the information environment so that the agents' predictable misinterpretation of the signals leads to a better collective outcome. The agents continue to act rationally given their beliefs. The beliefs are wrong in a direction the designer chose.

The through-claim is about the target of intervention. Incentive design assumes the agent's model is fixed and changes the world to redirect behavior. Cognitive arbitrage assumes the world is fixed and shapes the agent's perception to redirect behavior. Neither changes what the agent wants. One changes what the agent gets; the other changes what the agent sees. The distinction matters because perception is cheaper to modify than reality, and because agents with misspecified models are already acting on wrong information — the designer is choosing which wrong information rather than introducing wrongness from scratch.