Run a bounded evaluation

Launch only the selected evaluation, then explain what its evidence supports. For voice, success requires actual voice conversations and inspected recordings; a saved resource or a completed run status is not sufficient.

Preflight and budget

Confirm the intended org, workspace and API environment from the customer's context; authentication alone does not identify the organization. Reuse their agent/persona/test set. Read the exact cases and metric definitions before launching. Use coval --agent agent doctor, resource context and --help. Do not print secrets or change the machine's global account configuration.