Evaluation and recovery

Evaluation modes

• auto: use an available evaluator when applicable.
• required: fail closed or request review when evaluation cannot establish success.
• off: skip control-plane evaluation; the application still owns domain validation.

What is evaluated today

Deterministic validators cover JSON and tool-call schema requirements. Generic text may be marked not_evaluated when no evaluator applies.

Bounded recovery

allow_escalation and max_attempts permit a higher-confidence eligible route to be tried after a failed result. Cost, time, policy, privacy, and provider boundaries still apply.

{"mc1":{"evaluation":"required","allow_escalation":true,"max_attempts":2}}