Runtime. Soliton.

Runtime

The model runs next to the validator.

Most of the delay in an on-chain agent is not thinking. It is waiting for an endpoint on the other side of the internet to answer. Both routes below carry the same decision.

Each dot is one hop. The upper route crosses the public internet twice before the model has seen anything.

Your keys never meet the model.

The model proposes a transaction. It does not sign one. A policy engine sits between the two, and it is the only thing that can reach a key.

Simulate first: Every transaction is replayed against the current state before anything is signed. There is no flag to turn this off.
Bounds you set: Maximum value moved, maximum slippage, allowed programs, allowed counterparties. Written once, checked on every call.
Keys in an enclave: Signing happens in hardware the model cannot read from. It receives a yes or a no, never the key.
A refusal is a valid answer: S1 Verify returns a decision, and declining to sign counts as a correct one.

Does it land inside the window?

Finality is a distribution, not a number. Drag the marker to set how long the whole round trip takes and read the probability it beats the block.1

Probability the agent decides before finality

84.0%

Drag anywhere on the plot, or use the arrow keys, to set how long your agent takes to decide. The curve is finality modelled as a lognormal with a 260 ms median and a 900 ms 95th percentile, fitted to the published mainnet range. It is a model, not a measurement.

Three places, close to the leaders.

Inference sits where the validators are: Frankfurt, Amsterdam, Tokyo. If you run your own validator, Runtime installs beside it instead.

Read the API docs

Every latency on this page is a specification. The public route is built from published figures for a frontier model and a public RPC endpoint; the colocated route is a budget for a system that does not exist yet. What is not done.