Skip to main content
Your endpoint earns a reputation from how it actually behaves. An off-chain scorer probes it, a small set of rules decides what counts against you, and the result is recorded on-chain. The design has one guardrail at its center: a bad answer from your endpoint is your fault, but a bad input from the buyer is not.

​
The scorer

An off-chain scorer probes each endpoint on three things:
  • Schema. Does the response match the endpoint’s declared schema?
  • Latency. Does it respond within its bounds?
  • Correctness. Is the answer actually right for the input?
The scorer runs continuously, not just at build time, so reputation tracks how the endpoint behaves in the wild.

​
Three-way classification

Every response is classified into one of three outcomes. This is the same classification the escrow response gate uses to decide whether to charge, and it is the wrongful-strike guard.
The response validates against the endpoint’s declared success schema. The buyer is charged, and this is a healthy call. No strike.
The response is a well-formed “your input was bad” answer that validates against the endpoint’s declared error schema. This is the endpoint working correctly on bad buyer input, so it is never a strike. Depending on the error policy the call is free or charged a small declared-error price.
The response is neither a valid success nor a valid declared error, or the handler timed out or threw. Nothing is charged, and a strike is recorded against the creator.
A declared_error is never a strike. If your endpoint cleanly rejects bad input, that is it working, not failing. Only malfunctions and timeouts count against you. See metering and classification.

​
Five consecutive strikes deactivate the endpoint

Strikes are counted consecutively. Five consecutive strike-worthy failures (malfunctions or timeouts) deactivate the endpoint. A single success resets the counter to zero, so an endpoint that has an occasional hiccup but mostly works is never at risk. When an endpoint is deactivated:
1

The on-chain resource is paused

The endpoint stops taking paid calls. Its listing reflects that it is no longer live.
2

A bond-slash review is raised

Deactivation raises a review that can slash the endpoint’s bond into the insurance pool. Slashing is a two-step process with a 1-day dispute window, so it is never instant. See Pricing and bonds.

​
Reputation is on-chain

Reputation is recorded on-chain through an ERC-8004 reputation registry. It is a health score, not money: it captures how reliably an endpoint behaves so agents can decide whether to trust it, separate from the payment path. Money moves through the escrow and the split; reputation moves through the registry.

Metering and classification

How a response is classified as success, declared error, or malfunction.

ERC-8004 identity

The on-chain identity and reputation registry behind every endpoint.