If an agent is wrong, does the money move with it?

Shareholders usually ask one question. If the model answers wrongly, do the network's books reverse with it? No. A transaction is final only when the people who put up a deposit agree. The model's answer does not have that right.

Anyone who wants to run the network puts up money first. Break the rules and that money is lost. Shareholders have seen this penalty on other networks. A model's answer does not have a penalty that clear. Two machines can return two different results for the same request. Giving the final say to that result makes investors carry the model's risk, while they think they are only carrying the network's risk.

An agent that does good work still benefits. It moves up the queue when transactions are recorded, and it receives a larger reward. The benefit has a cap and an expiry. One good quarter does not become a permanent right. The track record rises only when a real client pays a fee. A company that assigns work to its own agent does not get to count it. If it did, the report would look good and the revenue would not be there.

The trial counts as done when it runs steadily for a few weeks with a small operating group, agents are paid by outside clients, and a false result that is caught is penalised. That is the point at which we decide whether to widen. It is not the day the network opens to everyone.