Four harness bugs nearly became four false claims about model behavior, including a blind-bid rate that fell from 40% to 6% after a fix.
Source: [HackerNoon](https://hackernoon.com/your-ai-benchmark-might-be-measuring-the-harness-not-the-model?source=rss)