Work is sent to several machines at once. They compute it separately, without being able to see each other's answers, and a majority has to agree before anybody is paid. A machine that returns something else loses part of its deposit.
Which does mean paying for the same work several times over. That is the cost of not having to trust any single one of them, and it is priced into what a job costs rather than hidden behind it.
How a job settles
Everything here rests on the machines not being able to see each other. Take that away and the rest of it stops meaning anything.
Payment is escrowed when the job is submitted, before anything is assigned. It is then sent to several independently owned machines at once, chosen on stake, on their record and on how busy they already are, rather than on whoever happens to be biggest.
The machines cannot see each other's answers. Each computes on its own and returns a hash of the result. If they could compare notes first, the majority would collapse into whoever answered fastest and the whole exercise would be theatre.
The hashes are compared. Where enough of them match, the job settles and everybody who agreed is paid. A machine returning something else is not paid and loses part of its deposit, so an answer somebody invented costs them money rather than earning it.
The threshold is a fraction rather than a fixed count, because how many machines a job goes to depends on what the network can spare. Ordinary work needs a simple majority. Work that can trigger a trade goes to more machines and needs two thirds.
Whose network this is
The mesh opens to other people's work at mainnet, not at some later phase once we have finished with it. Anyone can submit inference work and pay for it, and over time operators can serve their own models rather than only the ones we trained.
That distinction is the difference between a network and an outsourcing arrangement. Our trading engine is the first paying customer and the reason there is any demand on day one, which is a real advantage for the operators who join early. It is a customer, though, not the owner.
The same machinery answers oracle requests for any contract on the chain, with no new trust assumption, because a price read by five independent machines that have to agree is exactly the structure already built for inference.
submitted payment escrowed up front, before assignment
assigned to several nodes at once, chosen on stake,
on their record, and on current load
computed independently. nodes cannot see each other
returned a hash of the result, nothing else
compared enough matching → settled, agreeing nodes paid
one node differs → not paid, loses 2% of stake
no majority → nobody slashed, job reissued,
the job itself is flaggedThat last line matters more than it looks. If a job is written so badly that honest machines legitimately disagree, punishing them would turn a bad job into a weapon against the people running it. So a split with no majority costs nobody anything, and the job goes back with a flag on it.