What 50k Runs of a 5-Line Eval Taught Us | Dark Hacker News