From MatMul to inference

On Amadeus the live work is seeded matrix multiplication. The matrices are not a free choice. They are derived from a seed the network can reconstruct. Other nodes check the result. Top-compute and vault-backed nodes vote. Finality lands in about half a second.

That is how the chain is secured. Miners are paid for that work. The same operations show up inside model inference, which is why the multiply is a serious primitive and not a costume. It is still a primitive. A checked grid is not a served model.

We asked which of two miners did useful work. Both multiply large grids. Both get paid if the network can check the work. Only one of them did it because an agent asked for the result.

The comments split, which is the right split. One reading says the network needed the work, so it was useful. The other says useful means someone was waiting for the answer. Both readings can be true of a security budget. Only the second reading is inference.

Inference is a request with a caller. A model, a schedule, a result an agent on Hub or Embed asked for, then a path back to that agent. The multiply is one step inside that, not the whole job.

We do not have that loop closed. A verified multiply can still be a multiply nobody requested. Routing the work the chain already pays for into a confidential request a live agent made is the step in front of us. Saying it is done would be a different product from the one that is shipping.

One side is a calculation the network can check. The other is an agent that asked for the result. We can count the first. We are not claiming the second is closed.

That is the step in front of the network. Not a costume on the multiply. A door from the work the chain already pays for to a request a live agent made.

Most useful-work networks stop at the primitive. The multiply secures the chain, the economics look better than empty hashes, and the story ends. That is a real improvement on a lottery. It is not yet an agent getting an answer.

Amadeus already has the other side of the building, which is why the gap is worth naming. Hub is where someone composes an agent. Embed is where that agent can sit in a screen that already exists. Settlement is how the agent can pay. None of that turns a random grid into a served model. It does mean the customer for the next step is not hypothetical.

Until the request has a door, useful work is still a security budget. Useful intelligence is the customer. We have the first. We are building the second.