The engine uses a strict metaprompt system instruction to restrict the model's output to a lowercase 'yes' or 'no', applied against a standardised prompt framing that asks whether the brand is recommended for a designated topic.