PromptEvalGate
The publish gate a suite defines (1.63.0): which expectation types and which must items must hold, and the floor and margin of the judged items, which gate only while the suite's judge is calibrated. An Agent applies it only once its owner names the suite (AgentEvaluationGate).
An object with 3 fields.
Fields
| Field | Type | Required | Description |
|---|---|---|---|
expectation_types | array of ConversationTestExpectationType | yes | The expectation types that gate: a failed expectation of one of them in any repetition fails the gate. Every type when omitted at create; empty, expectations do not gate. |
must_floor | number | yes | The least share of a must item's applicable repetitions judged pass, per scenario and item, on the candidate arm; 0.8 when omitted at create (a target), 0 checks no floor. |
must_margin | number | yes | The most a must item's share may fall below the baseline arm's; 0.2 when omitted at create (a target), 1 checks no margin. |