Many maps.
One weather.

Make them agree. Then make them wrong.

Watch source-aware gossip beside echoed consensus. Choose “Shared delusion” to bias a whole region’s sensors. Then break the radio. Compare useful trips with forecast error: agreement alone does not deliver a parcel.

Can agents agree enough to act without agreeing confidently on a mistake?

What the little world actually does.

Couriers predict the risk of their next trip from local readings and delivered messages. Evidence carries source, sensing time, variance and expiry. The naive treatment repackages posteriors as fresh evidence; source-aware gossip retains origins. A central collector must receive reports and send its estimates back through fallible links.

The colored field is the evaluator’s view. Couriers do not receive it. Their world model is tested through an action and a prediction of its outcome, so a tidy shared database cannot count as intelligence by itself.

传感器带噪声，消息会延迟，多个个体也可能共享同一种误差。这里同时测量任务结果、预测误差与错误共识。图中的全局风险属于评估器，行动者只能使用实际得到的观测与消息。

Keep the inconvenient outcomes.

Provenance did not establish an overall useful-work advantage over echoed consensus. It improved forecast quality with stale maps, but a shared sensor bias could still fool it. The declared superiority criterion did not pass. Missing consensus trials remain missing, with the valid denominator retained.

Source-aware, expiring evidence should resist repeated rumors, but cannot repair a common sensor failure or manufacture observations across a partition.

Paired seeded simulations. Every retained result can be regenerated from the versioned source; these are synthetic experiments, not field validation.

One chosen seed is an illustration. The downloadable batch retains every world and treatment; null means unavailable. A confidence interval describes the sampled synthetic worlds, not the real world.