  n  filter      verbatim        q=0.05         q=0.1         q=0.2         q=0.3   clean items scoring > 0.5
  3      no         0.996         0.995         0.995         0.967         0.881                       0.795
  3     yes         0.994         0.997         0.998         0.991         0.936                       0.000
  5      no         0.994         1.000         0.997         0.993         0.953                       0.000
  5     yes         0.995         1.000         0.997         0.996         0.960                       0.000
  8      no         0.995         1.000         0.997         0.976         0.925                       0.000
  8     yes         0.995         1.000         0.998         0.982         0.928                       0.000
 13      no         0.995         1.000         0.990         0.896         0.822                       0.000
 13     yes         0.995         1.000         0.990         0.914         0.831                       0.000

model that memorized the corpus: accuracy 0.4 on unseen items, 0.98 on verbatim items, 0.40 + 0.58 (1-q) on paraphrases with rate q
  threshold 0.038; flagged: verbatim 0.99, paraphrase q=0.05: 1.00, q=0.1: 0.99, q=0.2: 0.91, q=0.3: 0.63, clean 0.000
  expected accuracy on all 1500 items 0.7103; on unflagged items 0.4455; true clean accuracy 0.4
