A small-world community where a rumor spreads, correction and goodwill accounts push back, and every exposure triggers a judgment call.
22 humans sit on a small-world ring, each watching a handful of neighbors. Rumor accounts push the rumor outward; correction and proactive accounts push back. Every message that lands is a judgment call: the receiving human may attend to it, let it shift their belief, and sometimes pass it on to their own neighbors. Once a human's accumulated exposure crosses a threshold, they flip to "accepted rumor" (or back to "rejecting"). The judgment engine you pick below decides how fast and how expensively those calls get made — that's what the service-time race at the bottom is comparing.
Same workload, two ways of serving it — every authored message and every judgment call, priced at the selected engine's latency versus an all-generative baseline. This mirrors the paper's projection: a fixed GLM workload of 7 authorship and 53 judgment calls falls from roughly 120–240s to 18–55s once judgments move to Jev.