The engine uses a strict binary prompt demanding a literal 'yes' or 'no' recommendation, counting an affirmative only when the model's output begins with 'yes' while treating negative or non-conforming responses as zero.