Failure → Hypothesis → Implement → Re-run same benchmark → Compare → 
Keep or reject

        __
        __