An agent or compute job that hands a program a result it can act on. Deterministic computation has a planned verification route. A commitment to a model and an output is not a proof of correct inference, and saying otherwise would be the easiest lie on this site.
Label
Concept
Job types
Compute planned, AiInference research
Nuvex path
No runtime
Deployments
None. There is nothing to run.
What the application needs
An agent that moves funds on-chain needs its off-chain step to be something more than a claim. The bar is the same as for randomness: a program has to be able to reject a result that was not produced the way it says it was.
- A statement of what was computed, over which input, by which code or model.
- Evidence a program can check, rather than a signature saying trust me.
- A cost that makes checking worth it on-chain.
- A defined failure mode when the work cannot be verified.
What Nuvex provides today
The job-agnostic request and the callback shape. Compute and AiInference are identifiers in the same enum as Vrf, and they share the request lifecycle, which is why adding a job family means adding a verifier and an off-chain executor rather than a new core program.
That is also the limit of it. On-chain programs do no GPU work, no HTTP and no model execution by design, and nothing outside the chain in this repository executes a job either.
- /healthok
- /metricsexposed
- VRF keynot loaded
- fulfilmentsnot submitted
What does not exist
- Any compute executor. ADR 0004 names deterministic agreement with commit and reveal at Milestone 6.
- Any inference runtime. The threat model's model-poisoning row records exactly that.
- Any proof of correct inference. A commitment to a model and an output is not one.
- Any measured route to ZK or zkML, which the ADR puts at Milestone 9 and only after cost is measured.
“AI: a commitment to model and output. That commitment is not a proof of correct inference.”
What a developer would have to build
Today, the whole thing, outside Nuvex. The useful question this concept answers is which parts would eventually be protocol work and which would stay the application's.
- The executor, its determinism guarantees, and the commitment format for its output.
- The dispute path, since every challenge, reject and fail edge in the request state machine stays forbidden under ADR 0001.
- The economics of paying for the work, which are unset while max_fee is stored and never charged.
- An honest product surface: ADR 0004 requires the inference milestone to say in the product what a commitment does not prove.
Honest status
Research is the weakest label on this site and it is the right one here. There is no runtime, no verifier and no benchmark, so this page is a description of an open problem that the protocol's own records already frame.
On this page
