First-hand information for everyone
This paper builds a reproducible computer benchmark that compares three ways of running an economy: a centralized computational planner, a decentralized agent-based market, and a hybrid “meta-market.” In the reported runs the planner produced lower average welfare losses than the other two systems. Across ten training worlds the planner’s mean loss was 0.1018, versus 0.3072 for the agent-based market and 0.2679 for the meta-market. Similar gaps appear in five nominal holdout worlds (planner 0.1100, market 0.3127, meta-market 0.2697).
The authors generate many synthetic “worlds.” Each world has 150 sectors, ten regions, seven social-need groups, and 140 firms. The model uses a Leontief input–output core. That is an accounting structure that ties each sector’s output to the inputs it needs from other sectors. The benchmark also includes capacity limits, different firm productivities, prices that adjust inside the simulation, a list of eleven welfare loss components (for example essential unmet demand, total unmet demand, inequality across groups, quality shortfalls, corruption, rights violations, logistics failures, innovation shortfalls, environmental burden, volatility, and administrative cost), randomized welfare weights, structural shocks, and adversarial stress tests.
The three allocation systems are simulated with distinct rules. The planner receives a noisy demand signal plus some local corrections, directly uses the production matrix to solve the Leontief system, enforces capacity and reserve rules, and prioritizes essential sectors. The agent-based market assigns firms to sectors, updates prices in response to excess demand over ten periods, and lets firms update output, quality, cash-mediated productivity, and reputation. The meta-market adds stronger arbitrage and hedging behavior on top of the market process. The benchmark also runs optimizing red-team adversaries and information-reporting experiments to probe failure modes.
Why this matters: the main contribution is methodological. The experiment shows how to place different institutional rules inside a single simulated environment, define outcomes ahead of time, and search for counterexamples. The implementation is reproducible. A local reproduction recovered the reported metrics exactly, and the numerical input–output core is stable (the paper reports a spectral-radius estimate near 0.55 and numerical residuals close to machine precision). In adversarial searches and in a stylized fuel-logistics shock the planner remained superior: for essential unmet demand the planner scored 0.0767, the market 0.7806, and the meta-market 0.6721.
Important caveats: the authors emphasize this is a methodological prototype and not proof that planning is empirically superior. Several design choices mechanically favor the planner. For example, the planner directly accesses the production matrix, capacities, essential-sector labels, and some technical coefficients while market firms do not. In a representative world 40.7% of sectors have no assigned firm. Rights-violation measures for the planner are fixed at zero, several audit flags are asserted rather than tested, the nominal holdout set largely reuses training world families, and the incentive experiment is not a formal Groves mechanism. Because of these asymmetries the results should be read as validation of a synthetic experimental architecture and a call for stronger tests. The paper closes with a concrete validation agenda: empirical calibration, structural holdouts, sensitivity analysis, uncertainty quantification, genuine mechanism-design tests, and independent red-team replication.