A peer-reviewed, pre-registered study of small multi-agent AI economies got one prediction exactly right and one demonstrably wrong.
Researchers ran two quantitative predictions against live economies of Claude Opus 4.8 agents trading in parimutuel-coupled markets, spending $138.76 in API costs on a fully reproducible, cached experiment. Result 1 held: relative wealth growth tracked relative claimed information to within 46 millinats, inside the pre-registered 50-millinat tolerance band. Coalition value behaved as expected — submodular when channels were independent, flipping supermodular when a designed XOR synergy was introduced. Result 2 did not hold: the mean-field residual-scaling law, which predicted a smooth population misalignment response to incentive and control levers, was falsified. Instead of a continuous curve, agents produced a step function across the dominance boundary, with bistable outcomes near that boundary determined by random seed rather than by lever strength.
The falsification matters because alignment researchers frequently invoke smooth mean-field models to argue that misalignment can be dialed down gradually as capability and incentive controls improve. This experiment found no tested LLM population, at any capability level, realizing the noise-maintained-dispersion regime that smooth model assumes — the underlying math breaks down at the very boundary where it would be most useful.
The methodology is the rare bright spot: everything was frozen in a public git chain before any run, every number re-derives mechanically from cached outputs, and the full protocol is re-runnable at zero cost. That standard of pre-registration is still unusual enough in AI research that the failure result is credible rather than convenient.