'"Looks fine" isn't evidence: Why 1 spot check hides silent LLM regression'

LLMs don't produce a single output. So why do we test them like they do? If you are shipping AI...