Developers
Start with behavior. Then integrate.
The Playground shows the difficult cases visually. The controlled Evaluation API exposes the same approved synthetic scenarios for developer evaluation.
Start here
Three steps.
See UNKNOWN, retry refusal, reconciliation, and persistent uncertainty before touching the API.
The API is controlled. Approved evaluators receive a base URL and Bearer credential.
Create a synthetic evaluation, request retry evaluation, reconcile, and inspect the public response model.
Scenarios
Three approved evaluation cases.
One harmless synthetic effect is durably applied, but the provider response disappears.
The provider applies one harmless synthetic effect and later readback confirms it.
The provider may have acted, but later observation still cannot establish enough evidence.
Developer model
Small surface. Explicit behavior.
The public API exposes evaluation behavior, not private engine architecture.
Evaluation access is controlled. The API endpoint is not published openly.
The surface does not accept arbitrary provider destinations or execution-authority fields.
Responses use SATISFIED, NOT_SATISFIED, or UNKNOWN.
Effect requests, durable effects, and provider lookups make the flagship behavior visible.
API reference
Inspect the evaluation contract.
Routes, authentication, request examples, retry behavior, reconciliation behavior, response shape, and limitations.