I-6 regression: hardened judges, 1 run each, held-out 1 + 2 (40 cases), via apply_judges
  advice           3/3
  clean            18/18
  out_of_scope     2/2
  promissory       2/2
  wrong_fact       15/15
  wrong facts marked SEND: 0   clean marked REVIEW: 0/18
