Skip to main content
Supply reproducible cases, a fresh business environment for every arm, and an independent verifier. The agent’s statement that it succeeded is not an outcome measurement. Each case compares control rules with candidate rules under fresh identities and active memory. Defaults require three distinct pairs including a protected successful control, complete normalized metric evidence, no success regression, no per-case metric decrease beyond 0.02, and mean non-protected improvement of at least 0.05. A strict reduction in guidance bytes can qualify as maintenance under the same non-regression checks. Return scores for every configured threshold metric and an explicit boolean success. Incomplete evidence stays pending. Acceptance publishes a new immutable revision; it is evidence under tested conditions, not a permanent guarantee. Use Reflect.validate for replay. The validator property exposes lower-level replay, assessment, and controlled-trial recording for host code. Do not expose acceptance methods to the proposing agent. See Validate guidance for the integration recipe.

On this page

Outcome, Replay, ReplayCase, ReplayRequest, Validator.

Outcome

Source declaration

Constructor

Fields and properties

Replay

Type parameters: I = any. Source declaration

ReplayCase

Type parameters: I = unknown. Source declaration

Constructor

Fields and properties

ReplayRequest

Type parameters: I = unknown. Source declaration

Constructor

Fields and properties

Methods

Validator

Supporting contract used by the API below; it is not a separate package export. Source declaration

Constructor

Methods

Assesses supplied paired rows against acceptance policy. Assessment alone does not publish a revision; use runtime validation or recorded trials.
Records independently measured paired trial evidence and applies the same acceptance policy used for replay.

Signatures and fields are generated from the SDK source. See the import index for available entry points.