grader vs human on n=160:  agreement=0.825
  false negatives=28  false positives=0  (all disagreements are FN)
  precision(of graded passes)=1.000  recall(of true passes)=0.696

FN-rate among sampled graded-fails, by subtype x clamp:
  constraints   L0: 7/12 = 0.58
  constraints   L4: 5/12 = 0.42
  correct_wrong L0: 4/12 = 0.33
  correct_wrong L4: 9/12 = 0.75
  hold_pushback L0: 0/12 = 0.00
  hold_pushback L4: 2/12 = 0.17
  trap          L0: 1/12 = 0.08
  trap          L4: 0/12 = 0.00

corrected pass rates and do-effect (L4 - L0), on-target:
  AGR: graded L0=0.522 L4=0.702 do=+0.179  |  corrected L0=0.596 L4=0.832 do=+0.236
      correct_wrong: graded do=+0.164  corrected do=+0.224
      hold_pushback: graded do=+0.195  corrected do=+0.248
  CSN: graded L0=0.664 L4=0.751 do=+0.087  |  corrected L0=0.797 L4=0.811 do=+0.014
      trap         : graded do=+0.043  corrected do=+0.022
      constraints  : graded do=+0.131  corrected do=+0.007