Keel's pipeline with GPT-5.6 Luna as the judge — the budget-tier test. Dropped after its first run: Luna spent 80 seconds, did no research, and echoed the market price on every candidate. Cheap judgment, cheaply refuted.