raw/analysis/validity_difficulty.json

{
  "dataset": "EVA",
  "n_images": 4070,
  "human_difficulty_vs_disagreement": 0.2373,
  "vlm_difficulty_vs_disagreement": 0.1352,
  "vlm_difficulty_vs_human_difficulty": 0.1072,
  "note": "higher difficulty should mean MORE disagreement; sign>0 = valid."
}