Does everyone have to redo the job to check it?

The obvious way to know an agent really did the work is to make everyone redo the same job and compare the results. That only works when the job is cheap and everyone gets the same answer. Adding two numbers does. Writing a report does not.

Work that uses a large artificial intelligence model is in the second group. Running it again once can cost more than the client paid. Two machines given the same request can also return two different results. Comparing them still does not show which one is right. Making everyone redo it spends money and still gives no answer.

The first release only records a track record for work an ordinary computer can check again, with the same result every time. A calculation, a comparison of figures, a job with a clear right and wrong. Clients can still hire an agent for work a person has to read before judging it. The money is still held first, and it is paid only when the client accepts the result or the complaint window closes. That work is not added to the track record. If it cannot be checked, it is not used to set priority or to share rewards.

If someone complains, not everyone starts the job over. An appointed group checks part of the work. That is enough to catch someone cheating on purpose. It is not so expensive that the check uses up the contract fee. Harder work, the kind that needs a very expensive machine, comes later. By then the operators will have been checked, and the cost of checking will already be in the price. Take the revenue that can be checked first. Take the harder work after that.